{
 "title": "The Claim Ledger, The Durability Curve",
 "url": "https://durabilitycurve.com/claims/",
 "license": "https://creativecommons.org/licenses/by/4.0/",
 "ledgers": [
  {
   "ledger": {
    "title": "Claim Ledger: AI Can Take the Practice Without Taking the Job",
    "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/",
    "essay": {
     "title": "AI Can Take the Practice Without Taking the Job",
     "url": "https://durabilitycurve.com/blog/the-reps-nobody-counts/",
     "published": "2026-10-02"
    },
    "last_checked": "2026-10-13",
    "format": "current",
    "counts": {
     "claims": 67,
     "by_status": {
      "VERIFIED": 63,
      "EXECUTED": 4
     },
     "verified_at_primary": 63,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-reps-nobody-counts/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Employment of 22 to 25s in highly AI-exposed jobs is about 19% below where it would be had it kept pace with less-exposed peers (June 2026)",
     "quote": "\"Employment among workers ages 22–25 in highly AI-exposed occupations now stands about 19% below where it would be if it had kept pace with employment among similarly aged workers in less-exposed occupations.\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12 (same figure in the paper, p.3: \"has since widened to 19% as of June 2026\")",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Our six main facts', second fact; paper p.3",
     "date_read": "2026-10-13"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same measure read 15% a year earlier (July 2025 vintage)",
     "quote": "\"the kept-pace shortfall was 15% at the July 2025 data vintage\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3",
     "date_read": "2026-10-01"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The 13% figure was an earlier regression estimate the study headlined, not the same measure (body: \"The 13% you may have seen\")",
     "quote": "\"Earlier versions of this study headlined regression estimates adjusting for firm-level shocks (a 13% relative decline as of July 2025 data; 16% as of September 2025 data).\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3",
     "date_read": "2026-10-01"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In levels, exposed 22 to 25s fell about 11% from Nov 2022 to June 2026 while the least-exposed grew about 10%",
     "quote": "\"employment of 22–25-year-olds in the two most exposed quintiles fell about 11% between November 2022 and June 2026, while employment of the same age group in the three least-exposed quintiles grew about 10%.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3",
     "date_read": "2026-10-01"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The data is ADP payroll records covering millions of US workers through June 2026",
     "quote": "\"high-frequency administrative payroll data from ADP, encompassing millions of workers through June 2026.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.2",
     "date_read": "2026-10-01"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Experienced workers show no comparable gap",
     "quote": "\"with no comparable gap for experienced workers\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3, fact (2); news post: 'Experienced workers show no comparable gap.'",
     "date_read": "2026-10-01"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The gap comes mainly through fewer hires of young workers, not more people leaving",
     "quote": "\"it operates primarily through reduced hiring rather than increased separations\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3, fact (4)",
     "date_read": "2026-10-01"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Declines concentrate where AI usage substitutes for the work; where it complements, employment is flat or rising",
     "quote": "\"declines are concentrated in occupations where AI usage substitutes for human tasks, while complementary usage is associated with flat or rising employment\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3, fact (5)",
     "date_read": "2026-10-01"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The automation link is strongest for the youngest (about -0.10 per SD for 22 to 25s) and shrinks with age",
     "quote": "\"For the youngest workers, only the automation coefficient is negative and statistically significant—about −0.10 per standard deviation for 22–25 year-olds—and it shrinks monotonically with age.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.19; Table 3 p.20",
     "date_read": "2026-10-01"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Jobs that rely on tacit knowledge show faster employment growth for experienced workers",
     "quote": "\"Occupations that involve tacit knowledge see faster employment growth for experienced workers.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3",
     "date_read": "2026-10-01"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The codified-knowledge result is not significant with an education control; the tacit result for experienced workers survives it",
     "quote": "\"The codified-knowledge gradient is not significant when controlling for college share quintile, demonstrating its overlap with education. The tacit-knowledge gradient for experienced workers, by contrast, survives this control.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.21, footnote 21",
     "date_read": "2026-10-01"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors say the findings are descriptive, not a causal estimate of AI's effect",
     "quote": "\"We caution that this work does not estimate a causal impact of AI: these are descriptive facts\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.4",
     "date_read": "2026-10-01"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The gap is larger in the ADP data than in national survey benchmarks",
     "quote": "\"the divergence is directionally consistent but more pronounced in the ADP panel than in national survey benchmarks\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.4",
     "date_read": "2026-10-01"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The divergence holds when tech firms and computer jobs are excluded and with interest-rate and remote-work controls",
     "quote": "\"The divergence remains when we exclude technology firms and computer occupations, when we control for exposure to interest-rate increases and remote work\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Is AI causing these changes?'",
     "date_read": "2026-10-01"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The gap kept widening after interest rates peaked",
     "quote": "\"The divergence has continued to widen through mid-2026, well after interest rates peaked.\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Is AI causing these changes?', first bullet",
     "date_read": "2026-10-01"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford's definition of tacit knowledge",
     "quote": "\"implicit, experience-based knowledge that is difficult to write down or teach in a classroom. It is acquired through practice, mentorship, hands-on experience, and repeated exposure to real situations.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "Appendix J, pp.133-134",
     "date_read": "2026-10-01"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford's definition of codified knowledge",
     "quote": "\"formal, standardized, documented knowledge that can be taught through education, textbooks, or written procedures. It is explicit and transferable through text.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "Appendix J, p.133",
     "date_read": "2026-10-01"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors' proposed mechanism: AI automates the checkable tasks that used to justify entry-level headcount",
     "quote": "\"AI may be automating the checkable, process-intensive tasks that historically justified entry-level headcount, while increasing the leverage of experienced staff\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.5",
     "date_read": "2026-10-01"
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford's codified/tacit scores were produced by a language model reading job titles and descriptions",
     "quote": "\"We query GPT-5.4-mini (OpenAI) with temperature = 0 to score each job independently on two dimensions\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "Appendix J, p.133",
     "date_read": "2026-10-01"
    },
    {
     "id": "20",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-20",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The automation-versus-complement split comes from Anthropic Economic Index usage data",
     "quote": "\"complementarity (“augmentative”), and overall-usage exposures from the Anthropic Economic Index (Handa et al., 2025)\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "Table 3 caption, p.20",
     "date_read": "2026-10-01"
    },
    {
     "id": "21",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide: automating entry-level work can lower growth and welfare even with no fall in entry-level employment",
     "quote": "\"improvements in entry-level automation increase output upon adoption but can reduce growth and welfare, even without reducing entry-level employment.\"",
     "removal_reason": "",
     "source": "Ide, Automation, AI, and the Intergenerational Transmission of Knowledge, arXiv 2507.16078v9 (2026-06-08)",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "22",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-22",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide's mechanism: novices get moved away from the most productive experts, so best practice spreads more slowly",
     "quote": "\"This occurs when such improvements reallocate novices away from the most productive experts, slowing the diffusion of best practices.\"",
     "removal_reason": "",
     "source": "Ide, Automation, AI, and the Intergenerational Transmission of Knowledge, arXiv 2507.16078v9 (2026-06-08)",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "23",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-23",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide's premise: novices acquire tacit knowledge by working alongside experts",
     "quote": "\"novices acquire tacit knowledge by working alongside experts\"",
     "removal_reason": "",
     "source": "Ide, Automation, AI, and the Intergenerational Transmission of Knowledge, arXiv 2507.16078v9 (2026-06-08)",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "24",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-24",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide's result comes from a theoretical model, not measured data",
     "quote": "\"I develop a task-based overlapping-generations model\"",
     "removal_reason": "",
     "source": "Ide, Automation, AI, and the Intergenerational Transmission of Knowledge, arXiv 2507.16078v9 (2026-06-08)",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "25",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-25",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Beane's study: a two-year ethnography at five sites plus interviews at 13 top US teaching hospitals",
     "quote": "\"a two-year, five-sited, comparative ethnographic study of learning in robotic and traditional surgical practice, and a blinded interview-based study of surgical learning practices at 13 top-tier teaching hospitals around the U.S.\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract (publisher record via Crossref)",
     "date_read": "2026-10-01"
    },
    {
     "id": "26",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-26",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In robotic surgery only a minority of trainees came to competence, through norm-challenging 'shadow learning'",
     "quote": "\"allowed only a minority of robotic surgical trainees to come to competence\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract",
     "date_read": "2026-10-01"
    },
    {
     "id": "27",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-27",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Robotic surgery greatly limited the trainee's role in the work",
     "quote": "\"the radically different practice of robotic surgery greatly limited trainees’ role in the work, making approved methods ineffective\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract",
     "date_read": "2026-10-01"
    },
    {
     "id": "28",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-28",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Outcome Beane names: a decreasing supply of experts relative to demand",
     "quote": "\"a decreasing supply of experts relative to demand\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract, final sentence",
     "date_read": "2026-10-01"
    },
    {
     "id": "29",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-29",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The three practices successful trainees used",
     "quote": "\"“premature specialization” in robotic surgical technique at the expense of generalist training; “abstract rehearsal” before and during their surgical rotations when concrete, empirically faithful rehearsal was prized; and “undersupervised struggle,” in which they performed robotic surgical work close to the edge of their capacity with little expert supervision—when norms and policy dictated such supervision\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract",
     "date_read": "2026-10-01"
    },
    {
     "id": "30",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-30",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The scene: the robot let the expert operate alone; residents removed smoke and fluids or watched from a second console",
     "quote": "\"The robotic surgical system allowed an expert to operate alone — the trainee was optional. Residents either held a laparoscopic tool to remove smoke and fluids from the patient, or watched the surgical action from a second trainee console.\"",
     "removal_reason": "",
     "source": "UC Santa Barbara news, 'Shining a Light on Shadow Learning' (2020-09-15), Beane interviewed",
     "source_url": "https://news.ucsb.edu/2020/020026/shining-light-shadow-learning",
     "locator": "paragraphs 4-5 (Beane quoted)",
     "date_read": "2026-10-01"
    },
    {
     "id": "31",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-31",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Beane's own estimate, in an interview: residents get ten to twenty times less practice",
     "quote": "\"“Residents get ten to twenty times less practice,” said Beane\"",
     "removal_reason": "",
     "source": "UC Santa Barbara news, 'Shining a Light on Shadow Learning' (2020-09-15), Beane interviewed",
     "source_url": "https://news.ucsb.edu/2020/020026/shining-light-shadow-learning",
     "locator": "paragraph 3 (interview estimate, not a figure from the ASQ paper; attribute as his estimate)",
     "date_read": "2026-10-01"
    },
    {
     "id": "32",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-32",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic's trial: programmers learning a new library with AI help learned less, and some ways of using it preserved learning",
     "quote": "\"We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average.\"",
     "removal_reason": "",
     "source": "Shen and Tamkin (Anthropic), How AI Impacts Skill Formation, arXiv 2601.20245 (2026-01-28)",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "abstract; second clause for the patterns: 'three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance' (used as a one-line link to our Learn with AI piece)",
     "date_read": "2026-10-01"
    },
    {
     "id": "33",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-33",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The test task lists are O*NET's published task statements for Market Research Analysts and Marketing Specialists and for Marketing Managers, every task on each page",
     "quote": "\"Market Research Analysts and Marketing Specialists\"; \"Marketing Managers\"",
     "removal_reason": "",
     "source": "O*NET OnLine (US Department of Labor), details pages, updated 2026",
     "source_url": "https://www.onetonline.org/link/details/13-1161.00",
     "locator": "Tasks table of each saved details page (primaries/onet-13-1161.00.html, onet-11-2021.00.html; both pages carry 'Updated 2026'); every row of each Tasks table extracted to primaries/onet-*-tasks.txt (13 and 21 rows).",
     "date_read": "2026-10-01"
    },
    {
     "id": "34",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-34",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Our test: with Stanford's definitions the sort gave identical calls on two reruns (0% flips) and separated the analyst list (54% codified) from the manager list (29%) by 25 points; the bare question flipped 5% to 15% of calls and agreed with the framework 95% of the time",
     "quote": "\"P1 separation (framework, A minus B): 25 points -> PASS\" / \"P2 stability (framework, worst list flip rate): 0% -> PASS\" / \"P3 framework vs control agreement: 95%\"",
     "removal_reason": "",
     "source": "First-party run, pre-registered and locked (lock 6c2e9e32760b, HFloyd97/durability-engine#905)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-10-01"
    },
    {
     "id": "35",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-35",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Residents still finished their residencies, licensed to use the robot",
     "quote": "\"They are finishing their residencies licensed to use robots in the operating room, but many are graduating without crucial surgical skills\"",
     "removal_reason": "",
     "source": "UC Santa Barbara news, 'Shining a Light on Shadow Learning' (2020-09-15), Beane interviewed",
     "source_url": "https://news.ucsb.edu/2020/020026/shining-light-shadow-learning",
     "locator": "paragraph 3 (Beane quoted)",
     "date_read": "2026-10-01"
    },
    {
     "id": "36",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-36",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Open surgery needed the trainee because it took four hands",
     "quote": "\"In traditional surgery, surgical trainees were literally required for the job because the open procedure required four hands and two people most of the time\"",
     "removal_reason": "",
     "source": "UC Santa Barbara news, 'Shining a Light on Shadow Learning' (2020-09-15), Beane interviewed",
     "source_url": "https://news.ucsb.edu/2020/020026/shining-light-shadow-learning",
     "locator": "paragraph 4 (Beane quoted)",
     "date_read": "2026-10-01"
    },
    {
     "id": "37",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-37",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Some of the divergence predates generative AI",
     "quote": "\"Some differential trends are visible before the widespread use of generative AI.\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Is AI causing these changes?', reasons for caution",
     "date_read": "2026-10-01"
    },
    {
     "id": "38",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-38",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide is an economist at IESE Business School; this version is dated June 2026",
     "quote": "\"Department of Economics, IESE Business School\" / \"June 9, 2026\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "title page, p.1 (author footnote; date line)",
     "date_read": "2026-10-01"
    },
    {
     "id": "39",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-39",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "O*NET's marketing manager task on pricing, and its market research analyst task on competitor prices",
     "quote": "\"Develop pricing strategies, balancing firm objectives and customer satisfaction.\" / \"Gather data on competitors and analyze their prices, sales, and method of marketing and distribution.\"",
     "removal_reason": "",
     "source": "O*NET OnLine (US Department of Labor), details pages, updated 2026",
     "source_url": "https://www.onetonline.org/link/details/11-2021.00",
     "locator": "Tasks tables (saved primaries/onet-*-tasks.txt)",
     "date_read": "2026-10-01"
    },
    {
     "id": "40",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-40",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Our second test: the feeder-tracing prompt was stable (0.86 overlap) and found all 3 expected links, but gave feeders for only 53% of the manager's calls against a 70% bar, so the kit ships that step as a hand-filled worksheet",
     "quote": "\"F1 stability (mean Jaccard): 0.86 -> PASS\" / \"F2 coverage (feeder in both runs): 53% -> FAIL\" / \"F3 face validity: 3/3 expected pairs in both runs -> PASS\"",
     "removal_reason": "",
     "source": "First-party run, pre-registered and locked (lock e33328c1c259, HFloyd97/durability-engine#906)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-10-01"
    },
    {
     "id": "41",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-41",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "O*NET's manager forecasting task and analyst trend-tracking task (worked example)",
     "quote": "\"Use sales forecasting or strategic planning to ensure the sale and profitability of products, lines, or services\" / \"Forecast and track marketing and sales trends, analyzing collected data.\"",
     "removal_reason": "",
     "source": "O*NET OnLine (US Department of Labor), details pages, updated 2026",
     "source_url": "https://www.onetonline.org/link/details/11-2021.00",
     "locator": "Tasks tables",
     "date_read": "2026-10-01"
    },
    {
     "id": "42",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-42",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In our feeder run, six of the manager's 15 tacit calls had no feeder in the analyst's work in either run (vendor negotiation, legal disputes, product display, three sustainability items); hiring and training staff got one in one run only",
     "quote": "\"No B task involves vendor negotiation or building distribution networks.\" / \"No B task touches legal, copyright, or royalty issues.\" (rows A10 and A11, marked none)",
     "removal_reason": "",
     "source": "First-party run (lock e33328c1c259)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-10-01"
    },
    {
     "id": "43",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-43",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Young workers' employment fell in occupations that involve codified knowledge",
     "quote": "\"We find employment declines for young workers in occupations that involve codified knowledge.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.3",
     "date_read": "2026-10-01"
    },
    {
     "id": "44",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-44",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Ide's model the best experts pass their expertise to several novices at once",
     "quote": "\"only the most knowledgeable experts manage production and transmit their expertise to multiple novices\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "45",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-45",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic's study was a randomised experiment on developers learning a new library",
     "quote": "\"We conduct randomized experiments to study how developers gained mastery of a new asynchronous programming library with and without the assistance of AI.\"",
     "removal_reason": "",
     "source": "Shen and Tamkin (Anthropic), How AI Impacts Skill Formation, arXiv 2601.20245 (2026-01-28)",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "abstract",
     "date_read": "2026-10-01"
    },
    {
     "id": "46",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-46",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors' hypothesis: AI substitutes more effectively for codified knowledge",
     "quote": "\"One possibility is that AI more effectively substitutes for knowledge that has been digitized and codified\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.5",
     "date_read": "2026-10-01"
    },
    {
     "id": "47",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-47",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide frames the concern: by letting senior workers do more on their own, AI-driven automation may cut entry-level opportunities and the labour-for-training exchange",
     "quote": "\"By allowing senior workers to accomplish more tasks independently, AI-driven automation may reduce entry-level opportunities and threaten the implicit contract in which younger workers exchange their labor for training and the prospect of future promotion\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "p.2, Introduction, second paragraph (line-wrapped \"se-nior\" in extraction)",
     "date_read": "2026-10-01"
    },
    {
     "id": "48",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-48",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In open surgery, trainees learned by doing more of each operation over time, using approved methods, and it worked",
     "quote": "\"I found that learning surgery through increasing participation using approved methods worked well in traditional (open) surgery, as current literature would predict.\"",
     "removal_reason": "",
     "source": "Beane, Shadow Learning, Administrative Science Quarterly 64(1):87-123 (online 2018-01-09)",
     "source_url": "https://doi.org/10.1177/0001839217751692",
     "locator": "abstract (publisher record via Crossref)",
     "date_read": "2026-10-01"
    },
    {
     "id": "49",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-49",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Education is the one control that shrinks the estimates, which may be an alternative explanation or the channel AI works through",
     "quote": "\"estimates attenuate when controlling for occupational education levels, which may reflect either an alternative explanation or the very channel through which AI operates\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "pp.3-4 (sentence crosses the page break; matches once the page number is removed); also p.22: 'Education is the one control that attenuates these estimates'",
     "date_read": "2026-10-01"
    },
    {
     "id": "50",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-50",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In occupations with a high share of college graduates the difference between more- and less-exposed jobs is muted",
     "quote": "\"muted differences between more-exposed and less-exposed occupations compared to our main results\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "Section 4, p.25 (occupations with a high share of college graduates)",
     "date_read": "2026-10-01"
    },
    {
     "id": "51",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-51",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide's slowdown does not require lost practice or lost jobs; a worse pool of mentors is enough",
     "quote": "\"this slowdown does not require an improvement in automation to reduce entry-level employment or deprive novices of hands-on practice; a compositional deterioration in the pool of mentors is sufficient\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "Introduction, p.4",
     "date_read": "2026-10-01"
    },
    {
     "id": "52",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-52",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide's own policy conclusion: look beyond entry-level headcount, because welfare can fall with the number of entry-level jobs unchanged",
     "quote": "\"policy assessment should look beyond aggregate entry-level employment: the deployment of new technologies may reduce welfare even if the total number of entry-level jobs remains unchanged, provided it reallocates novices away from the most skilled experts\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "Conclusion (line-end hyphenation 'wel-fare' in extraction; verbatim otherwise)",
     "date_read": "2026-10-01"
    },
    {
     "id": "53",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-53",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford's authors cite Ide for the entry-level mechanism",
     "quote": "\"while increasing the leverage of experienced staff (Ide, 2025; Garicano and Rayo, 2025)\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.5",
     "date_read": "2026-10-01"
    },
    {
     "id": "54",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-54",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a field study of 5,172 support agents, AI help improved less experienced workers' speed and quality and facilitated learning",
     "quote": "\"using data from 5,172 customer support agents\" / \"Less experienced and lower-skilled workers improve both the speed and quality of their output\" / \"AI assistance facilitates worker learning\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Li, Raymond, Generative AI at Work, arXiv 2304.11771 (NBER w31161)",
     "source_url": "https://arxiv.org/abs/2304.11771",
     "locator": "abstract (all three phrases)",
     "date_read": "2026-10-01"
    },
    {
     "id": "55",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-55",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford treats experienced workers as a comparison group",
     "quote": "\"consistent with the placebo role of experienced workers\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.12",
     "date_read": "2026-10-01"
    },
    {
     "id": "56",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-56",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ide: technology that puts more novices beside the most productive experts strengthens knowledge transmission and raises growth",
     "quote": "\"By contrast, technological improvements that increase the number of novices learning from the most productive experts strengthen knowledge transmission and raise growth.\"",
     "removal_reason": "",
     "source": "Ide, arXiv 2507.16078v9",
     "source_url": "https://arxiv.org/abs/2507.16078",
     "locator": "abstract, p.1",
     "date_read": "2026-10-01"
    },
    {
     "id": "57",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-57",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stanford measures tacit reliance partly through required experience and on-the-job training, and calls the codified/tacit estimates non-causal and purely suggestive",
     "quote": "\"We proxy reliance on tacit knowledge with required work experience and on-the-job training\" / \"We caution that these estimates are non-causal and purely suggestive\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Chandar, Chen, Canaries in the Coal Mine? (Aug 2026 revision), Stanford DEL",
     "source_url": "https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf",
     "locator": "p.19 (second sentence crosses the p.19/20 break)",
     "date_read": "2026-10-01"
    },
    {
     "id": "58",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-58",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "By November 2022 the exposed young workers' position had returned to roughly its pre-pandemic level, so the later decline is new",
     "quote": "\"By November 2022, the relative position of exposed young workers had already returned to roughly its pre-pandemic level.\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Is AI causing these changes?', reasons AI may be playing a role",
     "date_read": "2026-10-01"
    },
    {
     "id": "59",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-59",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Beane: residents lacked skills because they got too little practice or mentorship; the console time they did get was lower quality because surgeons 'helicopter taught' and retook control",
     "quote": "\"because they did not receive enough practice or mentorship\" / \"surgeons “helicopter taught” — giving frequent and very public feedback to residents at the console and periodically retaking control of the robot from them\"",
     "removal_reason": "",
     "source": "UC Santa Barbara news, 'Shining a Light on Shadow Learning' (2020-09-15), Beane interviewed",
     "source_url": "https://news.ucsb.edu/2020/020026/shining-light-shadow-learning",
     "locator": "paragraphs 3 and 5",
     "date_read": "2026-10-01"
    },
    {
     "id": "60",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-60",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Our sort marked the manager's pricing task tacit in both reruns",
     "quote": "\"Balancing objectives and satisfaction requires nuanced judgment from experience.\"",
     "removal_reason": "",
     "source": "First-party run (lock 6c2e9e32760b)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-10-01"
    },
    {
     "id": "61",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-61",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Of the 5,172 agents in the data, 1,636 are observed after getting the AI assistant",
     "quote": "\"In total, we observe the conversation text and outcomes associated with 3 million chats by 5,172 agents. Within this, we observe 1.2 million chats by 1,636 agents in the post-AI period.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Li, Raymond, Generative AI at Work, arXiv 2304.11771 (NBER w31161)",
     "source_url": "https://arxiv.org/abs/2304.11771",
     "locator": "Section 3 (data); Table 1 'Treated, Post' agents 1,636 (full PDF, primaries/blr-2304.11771.pdf)",
     "date_read": "2026-10-01"
    },
    {
     "id": "62",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-62",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The learning result is stated as evidence, not a firm finding",
     "quote": "\"We also find evidence that AI assistance facilitates worker learning\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Li, Raymond, Generative AI at Work, arXiv 2304.11771 (NBER w31161)",
     "source_url": "https://arxiv.org/abs/2304.11771",
     "locator": "abstract; full paper: the durable-learning evidence uses outage periods and the authors call those estimates noisy (Section 5)",
     "date_read": "2026-10-01"
    },
    {
     "id": "63",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-63",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The trial had 52 participants and was run through the Anthropic Fellows Program",
     "quote": "\"In our main study, 52 participants completed the task, 26 for each of the control and treatment groups.\" / \"Work done as a part of the Anthropic Fellows Program\"",
     "removal_reason": "",
     "source": "Shen and Tamkin (Anthropic), How AI Impacts Skill Formation, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "Section 4 (study design); title-page footnote (full PDF, primaries/anthropic-2601.20245.pdf)",
     "date_read": "2026-10-01"
    },
    {
     "id": "64",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-64",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Estimates adjusting for firms' overall hiring point the same way but shift with modelling choices",
     "quote": "\"estimates that account for overall changes in firm hiring are directionally consistent while becoming more sensitive to specification choices\"",
     "removal_reason": "",
     "source": "Stanford DEL news post, 2026-08-12",
     "source_url": "https://digitaleconomy.stanford.edu/news/canariesaug26/",
     "locator": "section 'Is AI causing these changes?', reasons for caution",
     "date_read": "2026-10-01"
    },
    {
     "id": "65",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-65",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Our earlier piece on reps for individual judgement",
     "quote": "\"It runs on reps, and every choice you hand over before trying it yourself is a rep you do not take.\"",
     "removal_reason": "",
     "source": "The Durability Curve, You Reach Before You Think (2026-08-12)",
     "source_url": "https://harryfloyd.substack.com/p/you-reach-before-you-think",
     "locator": "body, section before 'Why the reps are the thing'",
     "date_read": "2026-10-01"
    },
    {
     "id": "66",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-66",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The support-study authors suggest the assistant may capture and spread the behaviour of the most productive agents",
     "quote": "\"This finding suggests that generative AI systems may be capable of capturing and disseminating the behaviors of the most productive agents.\"",
     "removal_reason": "",
     "source": "Brynjolfsson, Li, Raymond, Generative AI at Work, arXiv 2304.11771 (NBER w31161)",
     "source_url": "https://arxiv.org/abs/2304.11771",
     "locator": "Introduction (full PDF, primaries/blr-2304.11771.pdf)",
     "date_read": "2026-10-01"
    },
    {
     "id": "67",
     "url": "https://durabilitycurve.com/claims/the-reps-nobody-counts/#c-67",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the trial, participants who fully delegated the coding got some productivity gain at the cost of learning the library",
     "quote": "\"Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library.\"",
     "removal_reason": "",
     "source": "Shen and Tamkin (Anthropic), How AI Impacts Skill Formation, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "abstract",
     "date_read": "2026-10-01"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Your AI Grader Is Only as Good as Its Answer Key",
    "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/",
    "essay": {
     "title": "Your AI Grader Is Only as Good as Its Answer Key",
     "url": "https://durabilitycurve.com/blog/your-ai-grader-answer-key/",
     "published": "2026-09-28"
    },
    "last_checked": "2026-09-27",
    "format": "current",
    "counts": {
     "claims": 68,
     "by_status": {
      "VERIFIED": 57,
      "EXECUTED": 11
     },
     "verified_at_primary": 57,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/your-ai-grader-answer-key/"
   },
   "rows": [
    {
     "id": "L1",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Without a correct reference, judges agreed well with human experts only on questions they had been able to answer themselves",
     "quote": "\"when not provided with a correct reference, judges show high agreement with human experts only on questions the judges were able to correctly answer themselves\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, abstract",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 1.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L1a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L1a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "GPT-4o as judge, Cohen's kappa with experts. Pairwise: questions it answered correctly 0.78 with no key, 0.92 with a human key; questions it got wrong 0.30 with no key, 0.16 with its own answer as the key, 0.83 with a human key. Single grading: questions it answered correctly 0.46 with no key and 0.59 with a human key; questions it got wrong 0.16 and 0.81",
     "quote": "\"GPT-4o Correct 0.78±0.02 0.86±0.02 0.92±0.02 0.46±0.07 0.52±0.06 0.59±0.06 Incorrect 0.30±0.09 0.16±0.08 0.83±0.06 0.16±0.14 0.13±0.14 0.81±0.13\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Table 4 (GPT-4o row)",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 21.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L1b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L1b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Pairwise, questions each judge got wrong, no key vs human key: Llama 3.3 70B 0.39 vs 0.92; Phi 4 -0.08 vs 0.75; Qwen 2.5 7B 0.21 vs 0.63; Yi 1.5 34B 0.13 vs 0.56 (own answer as key: 0.35, -0.08, 0.14, 0.02)",
     "quote": "\"Llama 3.3 70b Correct 0.79±0.03 0.86±0.03 0.93±0.02 0.61±0.08 0.54±0.08 0.73±0.07 Incorrect 0.39±0.06 0.35±0.07 0.92±0.03 0.13±0.06 0.23±0.10 0.69±0.09 Phi 4 Correct 0.67±0.07 0.88±0.04 0.91±0.04 0.48±0.09 0.47±0.07 0.67±0.07 Incorrect -0.08±0.09 -0.08±0.10 0.75±0.06 0.03±0.05 0.02±0.12 0.56±0.12 Qwen 2.5 7B Correct 0.65±0.05 0.69±0.04 0.78±0.04 0.31±0.11 0.37±0.09 0.54±0.09 Incorrect 0.21±0.07 0.14±0.07 0.63±0.06 0.08±0.07 0.21±0.10 0.56±0.08 Yi 1.5 34B Correct 0.31±0.12 0.53±0.12 0.77±0.09 0.23±0.11 0.29±0.08 0.43±0.09 Incorrect 0.13±0.09 0.02±0.09 0.56±0.08\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Table 4 (other judges)",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 21.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L1c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L1c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The paper notes that a relatively small model (Qwen 2.5 7B) with a human reference can grade better than a larger model (GPT-4o) without one, on the same set of responses",
     "quote": "\"providing a human reference to a relatively small model (such as Qwen 2.5 7B ) can yield better judgments than using a larger model without human references (such asGPT-4o) on the same set of responses\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §6",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 22.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L2",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "With its own answer as the key, GPT-4o's pairwise kappa was 0.86 on questions it had answered correctly and 0.16 on questions it could not answer",
     "quote": "\"decreases from 0.86 to 0.16 in the pairwise case on questions GPT-4o could not answer\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §6",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 23.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L3",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Averaged across all five graders on questions GPT-4o answered correctly (Table 3 does not say whether single and pairwise grading are pooled): kappa with a human key 0.69, a verified GPT-4o answer as key 0.61, a completely unrelated key 0.50, no key 0.46, a subtly wrong key 0.21",
     "quote": "\"Human 0.69 ±0.06 0.74±0.06 0.57±0.08 GPT-4o (✓) 0.61 ±0.07 0.66±0.07 0.52±0.09 None 0.46 ±0.08 0.56±0.08 0.22±0.09 Random 0.50 ±0.08 0.47±0.08 0.09±0.1 Wrong 0.21 ±0.06\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Table 3",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 19.3%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L3c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L3c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Table 3's figures are across all five graders",
     "quote": "\"Table 3 presents results for these reference types across all five judges\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §6",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 26.3%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L3a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L3a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The paper says a slightly incorrect reference can in some cases be worse than no reference at all",
     "quote": "\"in some cases be worse than\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §6",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 25.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L3d",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L3d",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors describe the Random key as a completely unrelated reference, and say a slightly incorrect reference can in some cases be worse than a completely unrelated one or none",
     "quote": "\"providing a slightly incorrect reference can in some cases be worse than providing a completely unrelated reference or no reference at all\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §6",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 25.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L3b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L3b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors conclude that verifying a model-generated answer, rather than writing and checking one by hand, was sufficient as a key in their setting",
     "quote": "\"Verifying responses generated by a model rather than manually writing and checking answers via annotators is sufficient in this setting\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, §7",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 28.6%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L4",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "2026 judges, questions each got wrong, against a stand-in label: single grading no key vs human key, Opus 4.7 0.26 vs 0.62, GPT-5.4 0.32 vs 0.67, Gemini 3.1 0.44 vs 0.71; pairwise 0.33 vs 0.66, 0.46 vs 0.87, 0.67 vs 0.95 (own answer as key, pairwise: 0.33, 0.45, 0.62)",
     "quote": "\"Opus 4.7 Correct 0.86±0.04 0.90±0.04 0.91±0.04 0.63±0.06 0.69±0.06 0.77±0.05 Incorrect 0.33±0.16 0.33±0.15 0.66±0.13 0.26±0.14 0.27±0.12 0.62±0.11 GPT-5.4 Correct 0.75±0.06 0.87±0.04 0.89±0.04 0.51±0.06 0.64±0.06 0.66±0.06 Incorrect 0.46±0.11 0.45±0.13 0.87±0.08 0.32±0.11 0.25±0.12 0.67±0.09 Gemini 3.1 Correct 0.78±0.06 0.80±0.06 0.83±0.05 0.55±0.06 0.62±0.06 0.69±0.06 Incorrect 0.67±0.13 0.62±0.14 0.95±0.07 0.44±0.13 0.35±0.14 0.71±0.10\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Appendix J, Table 12",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 90.1%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L4a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L4a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors treat the 2026-judge rows, scored against a stand-in label, as corroboration",
     "quote": "\"we regard these results as corroboration rather than equivalent to our main table\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Appendix J",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 88.7%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L4b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L4b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The stand-in label is GPT-4o grading with the human-verified references",
     "quote": "\"We used the GPT-4o Judge with the human-verified references as a proxy correctness label\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Appendix J",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 88.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L4c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L4c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors report that higher reasoning effort can narrow the no-reference gap, for example GPT-5.4 at high effort",
     "quote": "\"higher reasoning effort can narrow the no-reference gap (e.g.,GPT-5.4 at high effort)\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Appendix J",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 88.9%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L4d",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L4d",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "GPT-5.4 at high reasoning effort, on questions it got wrong, against the stand-in label: pairwise 0.68 with no key and 0.84 with the human key; single grading 0.38 and 0.67",
     "quote": "\"GPT-5.4 low Correct 0.85±0.05 0.90±0.04 0.54±0.06 0.57±0.05 Incorrect 0.55±0.13 0.95±0.06 0.34±0.11 0.56±0.09 default Correct 0.75±0.06 0.89±0.04 0.51±0.06 0.66±0.06 Incorrect 0.46±0.11 0.87±0.08 0.32±0.11 0.67±0.09 high Correct 0.80±0.06 0.87±0.05 0.68±0.08 0.67±0.06 Incorrect 0.68±0.14 0.84±0.11 0.38±0.19 0.67±0.11\"",
     "removal_reason": "",
     "source": "Krumdick et al., No Free Labels (COLM 2026), arXiv 2503.05061v3, Appendix J, Table 13 (GPT-5.4 rows)",
     "source_url": "https://arxiv.org/pdf/2503.05061v3",
     "locator": "page-text offset 91.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L5",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On MATH, DeepSeek V3 as a judge without a reference reached Scott's pi 0.72 against ground truth; as a matcher with the reference, 0.98",
     "quote": "\"DeepSeek v3 model achieves only modest agreement π = 0.72, while as a matcher, it achieves π = 0.98\"",
     "removal_reason": "",
     "source": "Chandak et al., Answer Matching Outperforms Multiple Choice (preprint), arXiv 2507.02856v1, §3.1",
     "source_url": "https://arxiv.org/pdf/2507.02856v1",
     "locator": "page-text offset 22.6%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L5a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L5a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A 1.7-billion-parameter Qwen3 model matching answers to the reference reached pi 0.97 against ground truth on MATH",
     "quote": "\"answer matching, even with the 1.7 billion parameter Qwen3 model (non-thinking mode), achieves near-perfect alignment with the ground-truth (π = 0.97)\"",
     "removal_reason": "",
     "source": "Chandak et al., Answer Matching Outperforms Multiple Choice (preprint), arXiv 2507.02856v1, §3.1",
     "source_url": "https://arxiv.org/pdf/2507.02856v1",
     "locator": "page-text offset 22.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L5b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L5b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Answer matching with recent models, even small ones, reached near-perfect agreement with human grading, in the range of inter-annotator agreement",
     "quote": "\"matching using recent models-even small ones-achieves near-perfect agreement, in the range of inter-annotator agreement\"",
     "removal_reason": "",
     "source": "Chandak et al., Answer Matching (preprint), arXiv 2507.02856v1, abstract",
     "source_url": "https://arxiv.org/pdf/2507.02856v1",
     "locator": "page-text offset 1.3%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L5c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L5c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On MATH, the 'true grades' come from MATH-Verify, a rule-based checker that compares each answer with the reference",
     "quote": "\"MATH-Verify library (Kydlicek et al., 2025) implements rule-based ground-truth evaluations of generative responses\"",
     "removal_reason": "",
     "source": "Chandak et al., Answer Matching (preprint), arXiv 2507.02856v1, §3.1",
     "source_url": "https://arxiv.org/pdf/2507.02856v1",
     "locator": "page-text offset 22.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L6",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "For frontier judges (DeepSeek V3, o4-mini), 80%+ of grading errors were false passes: the judge marked correct what human annotation marked incorrect",
     "quote": "\"errors disproportionately (80%+) arise from false positives\"",
     "removal_reason": "",
     "source": "Chandak et al., arXiv 2507.02856v1, §3.2",
     "source_url": "https://arxiv.org/pdf/2507.02856v1",
     "locator": "page-text offset 28.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L7",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On MT-Bench, without ties, GPT-4 agreed with expert labellers 85% of the time against 81% between humans",
     "quote": "\"The agreement under setup S2 (w/o tie) between GPT-4 and humans reaches 85%, which is even higher than the agreement among humans (81%).\"",
     "removal_reason": "",
     "source": "Zheng et al., Judging LLM-as-a-Judge (NeurIPS 2023 D&B), arXiv 2306.05685v4, §4.2 + Table 5",
     "source_url": "https://arxiv.org/pdf/2306.05685v4",
     "locator": "page-text offset 30.7%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L7a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L7a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "First turn, with ties counted (setup S1): GPT-4 grading pairs agreed with humans 66%, GPT-4 grading single answers 60%, humans with each other 63%; without ties (S2) both GPT-4 modes 85% against 81%",
     "quote": "\"G4-Pair 70% 1138 66% 1343 97% 662 85% 859 G4-Single - 60% 1280 - 85% 739 Human - 63% 721 - 81% 479 (a) First Turn\"",
     "removal_reason": "",
     "source": "Zheng et al. (NeurIPS 2023 D&B), arXiv 2306.05685v4, Table 5(a)",
     "source_url": "https://arxiv.org/pdf/2306.05685v4",
     "locator": "page-text offset 34.1%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L7b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L7b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In 2023 the MT-Bench authors proposed a reference-guided grader that first answers the question itself and uses its own answer as the reference; on their maths questions it cut the failure rate from 70% to 15%",
     "quote": "\"we first generate LLM judge's answer independently, and then display it as a reference answer in the judge prompt. In Table 4, we see a significant improvement in failure rate (from 70% to 15%) over the default prompt\"",
     "removal_reason": "",
     "source": "Zheng et al. (NeurIPS 2023 D&B), arXiv 2306.05685v4, §3.4",
     "source_url": "https://arxiv.org/pdf/2306.05685v4",
     "locator": "page-text offset 27.1%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L7c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L7c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The MT-Bench authors saw GPT-4 misjudge an answer to a maths problem it could solve when asked separately: it was misled by the answers it was shown",
     "quote": "\"although GPT-4 can solve the problem (when asked separately), it was misled by the provided answers, ultimately resulting in incorrect judgment\"",
     "removal_reason": "",
     "source": "Zheng et al. (NeurIPS 2023 D&B), arXiv 2306.05685v4, §3.3",
     "source_url": "https://arxiv.org/pdf/2306.05685v4",
     "locator": "page-text offset 24.1%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L26",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L26",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In 2023 the Prometheus authors found that removing the reference answer hurt their grader most, and said a reference relieves the grader of solving the question itself",
     "quote": "\"relieves the need for the evaluator LM to internally solve the instruction and only focus on assessing the response\"",
     "removal_reason": "",
     "source": "Kim et al., Prometheus (ICLR 2024, per the arXiv comments field), arXiv 2310.08491",
     "source_url": "https://arxiv.org/pdf/2310.08491",
     "locator": "page-text offset 45.7%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L8",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Re-tested on expert labels with a leave-one-annotator-out test, no LLM grader passed on MT-Bench, one of two datasets where none did",
     "quote": "\"in two datasets (MT-Bench, and SummEval), none of the LLMs pass the test\"",
     "removal_reason": "",
     "source": "Calderon, Reichart, Dror, The Alternative Annotator Test (ACL 2025), arXiv 2501.10970v4, §5",
     "source_url": "https://arxiv.org/pdf/2501.10970v4",
     "locator": "page-text offset 21.8%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L9",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Chatbot Arena's expert relabel, GPT-4 agreed with the two experts 81.0% and 78.5% (Llama-2-13b battles) and 76.3% and 79.3% (GPT-3.5-Turbo battles); the experts agreed with each other 89.8% and 79.4%",
     "quote": "\"Llama-2-13b Expert 1 Expert 2 GPT-4 Crowd 72.8% 77.8% 75.6% Expert 1 - 89.8% 81.0% Expert 2 - - 78.5% GPT-3.5-Turbo Expert 1 Expert 2 GPT-4 Crowd 73.8% 83.1% 75.6% Expert 1 - 79.4% 76.3% Expert 2 - - 79.3%\"",
     "removal_reason": "",
     "source": "Chiang et al., Chatbot Arena (ICML 2024), arXiv 2403.04132v1, Table 3",
     "source_url": "https://arxiv.org/pdf/2403.04132v1",
     "locator": "page-text offset 35.9%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L10",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On literature-grounded science answers, the best judge, o3, reached 65.1% accuracy",
     "quote": "\"Even the best-performing model, o3, achieves only 65.1% accuracy.\"",
     "removal_reason": "",
     "source": "Zhao et al., SciArena (NeurIPS 2025 D&B), arXiv 2507.01001v2, §6.2 + Table 1",
     "source_url": "https://arxiv.org/pdf/2507.01001v2",
     "locator": "page-text offset 32.6%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L10a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L10a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Random guessing on the same task scores 50.0",
     "quote": "\"Random Guess 50.0 o3 65.1\"",
     "removal_reason": "",
     "source": "Zhao et al., SciArena (NeurIPS 2025 D&B), arXiv 2507.01001v2, Table 3",
     "source_url": "https://arxiv.org/pdf/2507.01001v2",
     "locator": "page-text offset 32.3%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L10b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L10b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Expert annotators' average inter-annotator agreement: accuracy 0.82, kappa 0.76",
     "quote": "\"Average 0.82 0.76 0.94 0.91\"",
     "removal_reason": "",
     "source": "Zhao et al., SciArena, arXiv 2507.01001v2, Table 1",
     "source_url": "https://arxiv.org/pdf/2507.01001v2",
     "locator": "page-text offset 21.9%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L11",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The best frontier judge reached physician agreement on 3 of 5 clinical grading tasks (NEJM Healer, BIDMC ER, Landmark) and fell short on two; no single model matched on all five",
     "quote": "\"No single base model matched physician agreement across all five tasks.\"",
     "removal_reason": "",
     "source": "Buckley et al., Scaling Clinical Judgment to Evaluate Medical AI (preprint), arXiv 2609.12822v1, Results",
     "source_url": "https://arxiv.org/pdf/2609.12822v1",
     "locator": "page-text offset 15.1% · numbers present in text: 82%, 77%, 91%, 88%, 68%, 67%, 87%, 92%, 84%, 95%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L11a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L11a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Best judge vs physician agreement per task: NEJM Healer 82% vs 77%, BIDMC ER 91% vs 88%, Landmark 68% vs 67%; short on NEJM CPCs 87% vs 92% and Grey Matters Management 84% vs 95%",
     "quote": "\"reached physician inter-rater agreement on the NEJM Healer (Claude, 82% vs. 77% for physicians), BIDMC ER (Claude, 91% vs. 88%), and Landmark Diagnostic Cases (Gemini, 68% vs. 67%), but fell short on the NEJM CPCs (GPT-5, 87% vs. 92%) and the Grey Matters Management Cases (Claude, 84% vs. 95%)\"",
     "removal_reason": "",
     "source": "Buckley et al., Scaling Clinical Judgment to Evaluate Medical AI (preprint), arXiv 2609.12822v1, Results",
     "source_url": "https://arxiv.org/pdf/2609.12822v1",
     "locator": "page-text offset 14.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L11b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L11b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the clinical study, some tasks score a diagnosis against the known answer (the Bond score) and others score diagnostic reasoning against a rubric",
     "quote": "\"A score of 5 indicates that the correct diagnosis is included in the differential\"",
     "removal_reason": "",
     "source": "Buckley et al., Scaling Clinical Judgment to Evaluate Medical AI (preprint), arXiv 2609.12822v1, Results, Methods, Task Rubrics",
     "source_url": "https://arxiv.org/pdf/2609.12822v1",
     "locator": "page-text offset 48.9%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L11c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L11c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors' own grader, PrecepTron, fine-tuned on a small number of physician-scored cases, matched or exceeded physician agreement on three of five tasks, so no single model, base or fine-tuned, matched on all five (the abstract's 'physician-level consistent scoring across tasks' is looser than the Results and must not be cited)",
     "quote": "\"PrecepTron approaches physician inter-rater agreement, matching or exceeding it on three of five tasks (NEJM CPCs, NEJM Healer, and BIDMC ER)\"",
     "removal_reason": "",
     "source": "Buckley et al., Scaling Clinical Judgment to Evaluate Medical AI (preprint), arXiv 2609.12822v1, Results (PrecepTron section)",
     "source_url": "https://arxiv.org/pdf/2609.12822v1",
     "locator": "Results",
     "date_read": "2026-09-27"
    },
    {
     "id": "L12",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The GPT-4.1 grader using physician-written criteria scored above the average physician in five of seven themes (vendor-authored)",
     "quote": "\"GPT-4.1 as a grader exceeds the random baseline for all themes as shown in Table 5. It exceeds the average physician score in five out of seven themes\"",
     "removal_reason": "",
     "source": "Arora et al. (OpenAI), HealthBench, arXiv 2505.08775v1, §8.1 + Table 5",
     "source_url": "https://arxiv.org/pdf/2505.08775v1",
     "locator": "page-text offset 42.6%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L13",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Using the expert rubric, five of seven graders passed between 38.0% and 45.7% of the proofs experts failed and wrongly failed between 5.0% and 12.3% of the proofs experts passed; the other two passed 63.6% and 74.8% of the failed proofs (false fails 3.8% and 2.5%)",
     "quote": "\"82.7% 43.0% 5.2% 82.0% 45.7% 5.0% 81.8% 38.0% 8.8% 80.9% 41.4% 8.7% 78.9% 39.6% 12.3% 77.0% 63.6% 3.8% 74.2% 74.8% 2.5%\"",
     "removal_reason": "",
     "source": "Gonzalez et al., QEDBench (ICML 2026), arXiv 2602.20629v3, Fig. 5",
     "source_url": "https://arxiv.org/pdf/2602.20629v3",
     "locator": "page-text offset 5.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L13a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L13a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The paper defines leniency as the false-positive rate and harshness as the false-negative rate against the expert rubric pass mark",
     "quote": "\"We decompose errors into Leniency Rate(False Positives) and Harshness Rate(False\"",
     "removal_reason": "",
     "source": "Gonzalez et al., QEDBench, Fig. 5 caption",
     "source_url": "https://arxiv.org/pdf/2602.20629v3",
     "locator": "page-text offset 5.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L13b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L13b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The paper reports significant positive bias for certain frontier graders it names (Claude Opus 4.5, DeepSeek-V3, Qwen 2.5 Max, Llama 4 Maverick); it does not say this of all graders",
     "quote": "\"frontier evaluators like Claude Opus 4.5, DeepSeek-V3, Qwen 2.5 Max, and Llama 4 Maverick exhibit significant positive bias\"",
     "removal_reason": "",
     "source": "Gonzalez et al., QEDBench, abstract",
     "source_url": "https://arxiv.org/pdf/2602.20629v3",
     "locator": "page-text offset 0.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L13c",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L13c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "QEDBench's leniency and harshness figures are for graders using the expert rubric",
     "quote": "\"Judge Reliability Metrics (Expert Rubric Only)\"",
     "removal_reason": "",
     "source": "Gonzalez et al., QEDBench (ICML 2026), arXiv 2602.20629v3, Fig. 5 label",
     "source_url": "https://arxiv.org/pdf/2602.20629v3",
     "locator": "page-text offset 5.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L14",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Order-swap flip rate (MT-Bench / JudgeBench): Gemini 3.1 Pro 0.035 / 0.020; Claude Opus 4.6 0.038 / 0.022",
     "quote": "\"Gemini 3.1 Pro 0.977 0.989 0.035 0.978 0.989 0.020 Claude Sonnet 4 0.960 0.976 0.074 0.941 0.966 0.146 GPT-4o-mini 0.959 0.975 0.127 0.896 0.939 0.380 Claude Opus 4.6 0.958 0.979 0.038 0.974 0.986 0.022\"",
     "removal_reason": "",
     "source": "Norman, Rivera, Hughes, Reliability without Validity (preprint), arXiv 2606.19544v1, App. Table 6",
     "source_url": "https://arxiv.org/pdf/2606.19544v1",
     "locator": "page-text offset 78.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L14a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L14a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Order-swap flip rate (MT-Bench / JudgeBench): GPT-5.4 0.114 / 0.105",
     "quote": "\"GPT-5.4 0.932 0.965 0.114 0.933 0.966 0.105\"",
     "removal_reason": "",
     "source": "Norman, Rivera, Hughes, Reliability without Validity (preprint), arXiv 2606.19544v1, App. Table 6",
     "source_url": "https://arxiv.org/pdf/2606.19544v1",
     "locator": "page-text offset 78.8%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L15",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "2026 JudgeBench, chance-corrected agreement (Cohen's kappa): Gemini 3.1 Pro 0.841, Claude Opus 4.6 0.875, Claude Sonnet 4.6 0.782, GPT-5.4 0.606; exact match 0.964, 0.956, 0.920, 0.812 (one order, ties excluded)",
     "quote": "\"Gemini 3.1 Pro 0.849 0.511 33.8 0.964 0.841 12.3 0.956 0.898 5.9 Claude Opus 4.6 0.848 0.489 35.9 0.956 0.875 8.1 0.943 0.879 6.4 DeepSeek V3.2 0.845 0.486 35.9 0.791 0.545 24.5 0.921 0.826 9.5 Claude Sonnet 4.6 0.851 0.484 36.7 0.920 0.782 13.8 0.942 0.871 7.1 Llama 3.3 70B 0.841 0.465 37.6 0.664 0.283 38.1 0.892 0.769 12.3 Kimi K2.5 0.846 0.461 38.5 0.864 0.720 14.5 0.937 0.873 6.4 GPT-5.4 0.836 0.457 38.0 0.812 0.606\"",
     "removal_reason": "",
     "source": "Norman, Rivera, Hughes, Reliability without Validity (preprint), arXiv 2606.19544v1, §4.1 Table 2",
     "source_url": "https://arxiv.org/pdf/2606.19544v1",
     "locator": "page-text offset 24.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L15b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L15b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same authors warn that raw agreement does not correct for chance and overstates how well a grader discriminates",
     "quote": "\"This family of metrics does not correct for chance\"",
     "removal_reason": "",
     "source": "Norman, Rivera, Hughes (preprint), arXiv 2606.19544v1, §2",
     "source_url": "https://arxiv.org/pdf/2606.19544v1",
     "locator": "page-text offset 4.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L16",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Human annotators preferred the markdown side 57% of the time; four of five judges preferred it 73%–97% on the same pairs",
     "quote": "\"human annotators prefer the markdown side only 57% of the time, while four of the five judges prefer it 73%-97% on the same pairs\"",
     "removal_reason": "",
     "source": "Soumik, bias mitigation in LLM-as-a-Judge (TMLR 06/2026), arXiv 2604.23178v2",
     "source_url": "https://arxiv.org/pdf/2604.23178v2",
     "locator": "page-text offset 44.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L16a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L16a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The human comparison was two independent annotators on a 30-pair subsample: they preferred markdown 57% of the time on average, while four of the five judges preferred it 73%-97%",
     "quote": "\"two independent annotators on a 30-pair subsample preferred markdown only 57% of the time on average, while four of the five judges preferred markdown 73%-97%\"",
     "removal_reason": "",
     "source": "Soumik, bias mitigation in LLM-as-a-Judge (TMLR 06/2026), arXiv 2604.23178v2, §1 / App. G",
     "source_url": "https://arxiv.org/pdf/2604.23178v2",
     "locator": "page-text offset 28.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L16b",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L16b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The style test compared markdown with plain prose",
     "quote": "\"STYLE (markdown vs. plain prose)\"",
     "removal_reason": "",
     "source": "Soumik, bias mitigation in LLM-as-a-Judge (TMLR 06/2026), arXiv 2604.23178v2, §3",
     "source_url": "https://arxiv.org/pdf/2604.23178v2",
     "locator": "page-text offset 22.0%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L17",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "One study finds most measured self-preference is judges being unsure on hard items (its own summary: 89.6%, substantially reducing but not eliminating the evidence)",
     "quote": "\"evaluator uncertainty accounts for an average of 89.6%\"",
     "removal_reason": "",
     "source": "Roytburg et al., Are LLM Evaluators Really Narcissists? (ICML 2026), arXiv 2601.22548v4, §1",
     "source_url": "https://arxiv.org/pdf/2601.22548v4",
     "locator": "page-text offset 9.2% · numbers present in text: 89.6%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L18",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "GPT-5 as judge: false-positive rate on its own wrong GSM8K answers 82.0 [72.4, 90.7], 67.3 points above its rate on others' wrong answers",
     "quote": "\"GPT-5 82.0[72.4, 90.7]+67.3\"",
     "removal_reason": "",
     "source": "Zhang et al. (Microsoft/MIT), Can We Trust LLM Judges (preprint), arXiv 2609.12002v1, App. B Table 10 (GSM8K)",
     "source_url": "https://arxiv.org/pdf/2609.12002v1",
     "locator": "page-text offset 87.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L19",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Across three QA sets (NQ, TQA, HPQA), the panel (PoLL) beat GPT-4 alone on all three (0.763 vs 0.627; 0.906 vs 0.841; 0.867 vs 0.830), edged its best member on two (0.763 vs 0.749; 0.906 vs 0.902) and fell below Haiku on the third (0.867 vs 0.873)",
     "quote": "\"GPT-4 0.627 0.841 0.830 CMD-R 0.734 0.902 0.815 Haiku 0.749 0.894 0.873 GPT-3.5 0.726 0.859 0.833 PoLL 0.763 0.906 0.867\"",
     "removal_reason": "",
     "source": "Verga et al. (Cohere), Replacing Judges with Juries, arXiv 2404.18796v2, Table 1",
     "source_url": "https://arxiv.org/pdf/2404.18796v2",
     "locator": "page-text offset 18.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L21",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "When the judge is no better than the model it grades, correcting it with human labels can at most halve the labels you need",
     "quote": "\"no debiasing method can decrease the required amount of ground truth labels by more than half\"",
     "removal_reason": "",
     "source": "Dorner, Nastl, Hardt (ICLR 2025 Oral), arXiv 2410.13341v4, abstract",
     "source_url": "https://arxiv.org/pdf/2410.13341v4",
     "locator": "page-text offset 1.4%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L22",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L22",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the original JudgeBench (2024), many strong graders, GPT-4o among them, scored only slightly better than random on hard correctness pairs",
     "quote": "\"performing just slightly better than random guessing\"",
     "removal_reason": "",
     "source": "Tan et al., JudgeBench (ICLR 2025), arXiv 2410.12784v2, abstract",
     "source_url": "https://arxiv.org/pdf/2410.12784v2",
     "locator": "page-text offset 1.8%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L22a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L22a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "JudgeBench found a grader's ability to verify answers highly correlated with its ability to solve the problem itself",
     "quote": "\"the ability of the judge to verify the solution pairs is highly correlated with its ability to solve the problem itself\"",
     "removal_reason": "",
     "source": "Tan et al., JudgeBench (ICLR 2025), arXiv 2410.12784v2, §4.4",
     "source_url": "https://arxiv.org/pdf/2410.12784v2",
     "locator": "page-text offset 41.3%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L23",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L23",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "PRIOR ART: Eugene Yan's August 2024 review of LLM-evaluators drew on about two dozen papers and already covers reference-based evaluation and position bias",
     "quote": "\"Drawing from two dozen papers\"",
     "removal_reason": "",
     "source": "Eugene Yan, Evaluating the Effectiveness of LLM-Evaluators (aka LLM-as-Judge), Aug 2024",
     "source_url": "https://eugeneyan.com/writing/llm-evaluators/",
     "locator": "page-text offset 1.5%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L24",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L24",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "PRIOR ART: Hamel Husain's guide (Oct 2024) has a principal domain expert make pass/fail judgments and tracks the judge's agreement with them; the warning that agreement misleads on imbalanced data was added in a 2025 revision",
     "quote": "\"I also tracked agreement rates over time to ensure we were converging on a good prompt\"",
     "removal_reason": "",
     "source": "Hamel Husain, Using LLM-as-a-Judge For Evaluation: A Complete Guide (Oct 2024)",
     "source_url": "https://hamel.dev/blog/posts/llm-judge/",
     "locator": "page-text offset 60.7%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L24a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L24a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Husain's guide, as revised in September 2026, advises about 100 examples per failure mode to validate an automated grader and says below 60 the confidence intervals are often too wide; the October 2024 original did not contain this advice",
     "quote": "\"Aim for about 100 examples per failure mode, with enough Pass and Fail examples to measure both classes. Below 60 examples, the confidence intervals are often too wide\"",
     "removal_reason": "",
     "source": "Hamel Husain, Using LLM-as-a-Judge For Evaluation (Oct 2024)",
     "source_url": "https://hamel.dev/blog/posts/llm-judge/",
     "locator": "page-text offset 50.8%",
     "date_read": "2026-09-26"
    },
    {
     "id": "L25",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L25",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In code review against a written requirement, GPT-4o wrongly failed correct code (false-negative rate) 26.2% (HumanEval) and 35.9% (MBPP) when judging directly, rising to 73.2% and 87.9% when also asked to explain and repair",
     "quote": "\"GPT-4o achieves a relatively low FNR in HumanEval (26.2%) and MBPP (35.9%), but once explana- tions and repairs are required, the FNR increases sharply to 73.2% in HumanEval and 87.9% in MBPP\"",
     "removal_reason": "",
     "source": "Jin & Chen, Are LLMs Reliable Code Reviewers? Systematic Overcorrection (Automated Software Engineering 33, art. 90, 26 Jun 2026; read as arXiv 2603.00539v1, and the published text prints the same sentence), §5.2",
     "source_url": "https://arxiv.org/pdf/2603.00539v1",
     "locator": "page-text offset 38.7% · numbers present in text: 26.2, 35.9, 73.2, 87.9",
     "date_read": "2026-09-26"
    },
    {
     "id": "L25a",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-L25a",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In that paper, FNR is rejecting correct code (a false fail) and FPR is accepting buggy code (a false pass)",
     "quote": "\"FNR reflects over-correction (rejecting correct implementations), while FPR reflects unsafe acceptance\"",
     "removal_reason": "",
     "source": "Jin & Chen, Are LLMs Reliable Code Reviewers? Systematic Overcorrection (Automated Software Engineering 33, art. 90, 2026; read as arXiv 2603.00539v1), Table 2 caption",
     "source_url": "https://arxiv.org/pdf/2603.00539v1",
     "locator": "page-text offset 37.2%",
     "date_read": "2026-09-26"
    },
    {
     "id": "P1",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-P1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The prior registered before any paper was read: graders would hold on easy, preference-style chat grading and fail on correctness of hard items they cannot solve themselves and in specialised domains",
     "quote": "judges HOLD on easy, preference-style chat grading and FAIL on (a) correctness of hard items the judge cannot itself solve and (b) specialised domains",
     "removal_reason": "",
     "source": "first-party: pre-registration, 19:57 BST 2026-09-26",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J1",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "As run, scored to the precision each question asked for (one pass, n = 24 arithmetic-style questions with one right answer each, listed in Line/judge-check/items.py, 2026-09-26 21:02 BST): Claude Haiku 4.5, answering without tools, got 4 of 24 wrong. Haiku 4.5 grading those answers without a key passed 3 of its 4 wrong answers; given the verified key, it passed 0 of 4",
     "quote": "HEADLINE haiku passed 3/4 of its own wrong answers without a key, 0/4 with one",
     "removal_reason": "",
     "source": "first-party: Judge Check run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J2",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Claude Sonnet 5 grading the same answers without a key agreed with the key on 23 of 24: no false passes, one false fail (a correct final answer reached by an unjustified leap, which it failed although told to grade the final answer); with the key, 24 of 24",
     "quote": "no key agree 23/24 · false passes 0 · false fails 1 · unparsed 0 / with key agree 24/24",
     "removal_reason": "",
     "source": "first-party: Judge Check run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J3",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J3",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Re-dressing the same answers in markdown changed 2 of 24 of Haiku's verdicts (both from pass to fail on wrong answers) and none of Sonnet's",
     "quote": "markdown re-dress changed 2/24 verdicts / markdown re-dress changed 0/24 verdicts",
     "removal_reason": "",
     "source": "first-party: Judge Check run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J4",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J4",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "As run (the precision each question asked for): Haiku as judge could not solve 2 of the 24 items itself and, without a key, passed the wrong answer on both; on H6 (digit sum of 50!) it passed a wrong answer although it solved the item correctly when asked. Under the registered rule it could not solve 1 item (H11) and passed the wrong answer on it; H6 holds under both rules",
     "quote": "solved it itself: 22/24 · no-key agree on solved 21/22 · on unsolved 0/2",
     "removal_reason": "",
     "source": "first-party: Judge Check run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J5",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J5",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The run cost $2.83. As run, predictions P1, P4 and P5 passed, P3 passed for Haiku, P2 and Sonnet's P3 were unscorable. Under the pre-registered scoring rule (score.py --registered), P1 failed for both graders, P5 failed, Sonnet's P3 failed, P4 passed, Haiku's P3 passed",
     "quote": "cost $2.8326 · Predictions (pre-registered)",
     "removal_reason": "",
     "source": "first-party: Judge Check run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "T1",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-T1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Of 19 studies we found that report which way AI graders err against human or ground-truth labels, 12 lean toward passing flawed work (4 of them weakly, adversarially or in one direction only), 3 toward failing sound work (code review against a spec; older-model essay grading; a frontier panel on clinical diagnoses) and 4 are mixed",
     "quote": "Tally (Part 2, excluding 7b): Supports 12 (1, 2, 3, 4, 5, 6, 9, 13, 14, 16, 18, 19; of these, 13, 16, 18 and 19 are weak, adversarial or one-directional) · Contradicts 3 (7, 8, 11) · Mixed 4 (10, 12, 15, 17)",
     "removal_reason": "",
     "source": "first-party: error-direction sweep, 2026-09-26",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "T2",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-T2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The evidence base is 77 distinct studies: 63 in the four sweeps and 14 more in the error-direction check",
     "quote": "PAPERS TOTAL 77 distinct studies",
     "removal_reason": "",
     "source": "first-party: paper count",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "J6",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-J6",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Under the pre-registered scoring rule (0.001 relative tolerance), Haiku's answers were wrong on 2 of 24; Haiku grading them passed both wrong answers without a key and neither with one; both graders, given the key, failed the two answers that missed the precision the question asked for",
     "quote": "HEADLINE haiku passed 2/2 of its own wrong answers without a key, 0/2 with one",
     "removal_reason": "",
     "source": "first-party: Judge Check run, rescored",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "W1",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-W1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Our earlier piece A Green Score Is Not Evidence reported a groundedness metric that scored its best with the evidence removed",
     "quote": "A groundedness metric scored its best with the evidence removed.",
     "removal_reason": "",
     "source": "first-party: published article (subtitle)",
     "source_url": "https://harryfloyd.substack.com/p/a-green-score-is-not-evidence",
     "locator": "",
     "date_read": "2026-09-26"
    },
    {
     "id": "W2",
     "url": "https://durabilitycurve.com/claims/your-ai-grader-answer-key/#c-W2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Our earlier piece Your AI Looks Best Where You Can Check It Least argued that the work that is hardest to check is where AI output looks best",
     "quote": "Your AI Looks Best Where You Can Check It Least",
     "removal_reason": "",
     "source": "first-party: published article (title)",
     "source_url": "https://harryfloyd.substack.com/p/your-ai-looks-best-where-you-check-least",
     "locator": "",
     "date_read": "2026-09-26"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Amazon Shut Out Meta's Agent. Shopify Gave It Shop Pay.",
    "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/",
    "essay": {
     "title": "Amazon Shut Out Meta's Agent. Shopify Gave It Shop Pay.",
     "url": "https://durabilitycurve.com/blog/amazon-shut-out-metas-agent/",
     "published": "2026-09-26"
    },
    "last_checked": "2026-10-07",
    "format": "current",
    "counts": {
     "claims": 33,
     "by_status": {
      "VERIFIED": 33
     },
     "verified_at_primary": 33,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/"
   },
   "rows": [
    {
     "id": "AM1",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "At its Connect event, Meta said Muse, its personal AI agent, can now operate any app on a person's Mac, with their permission.",
     "quote": "With your permission, Muse can now drive any app on your Mac.",
     "removal_reason": "",
     "source": "Meta, \"Everything We Announced at Meta Connect 2026\", 23 September 2026",
     "source_url": "https://www.meta.com/blog/meta-connect-2026-everything-we-announced/",
     "locator": "§ Take Your Muse Everywhere You Go, item \"Muse for Mac update\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM2",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Meta said Muse will get its own email address and will come to Meta's AI glasses in the coming months.",
     "quote": "In the coming months we're bringing Muse to our AI glasses … Muse will have its own email address",
     "removal_reason": "",
     "source": "Meta, \"Everything We Announced at Meta Connect 2026\", 23 September 2026",
     "source_url": "https://www.meta.com/blog/meta-connect-2026-everything-we-announced/",
     "locator": "§ Take Your Muse Everywhere You Go, items \"Muse on AI glasses\" and \"Muse gets its own email address\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM3",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Meta said Muse launched with access to the entire Shopify catalogue, and that it is adding retailers including Walmart and Best Buy, plus Shop Pay and PayPal for payments.",
     "quote": "We launched with … access to the entire Shopify catalogue. … we’re adding Walmart, Best Buy, … plus Shop Pay and PayPal for payments.",
     "removal_reason": "",
     "source": "Meta Newsroom, \"The Biggest News From Connect 2026\", 24 September 2026",
     "source_url": "https://about.fb.com/news/2026/09/the-biggest-news-from-connect-2026/",
     "locator": "Item \"More connectors for Muse\" (the same paragraph appears on Meta's Connect blog post of 23 September)",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM4",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Meta says credentials a person shares with Muse go into secure storage, which Muse can use without seeing them.",
     "quote": "Any credentials a person shares with Muse go into secure storage, so Muse can use them without seeing them.",
     "removal_reason": "",
     "source": "Meta Newsroom, \"Introducing Muse: The World's First Personal AI Agent Built for Everyone\", 8 September 2026 (page updated 30 September 2026)",
     "source_url": "https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/",
     "locator": "§ How It Works, paragraph beginning \"Muse has no visibility\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM5",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Shopify's chief executive, Tobi Lütke, announced a partnership with Muse to enable checkout with Shop Pay on all Shopify stores.",
     "quote": "We are excited to announce we are partnering deeply with Muse to enable agentic checkout with Shop Pay on all Shopify stores",
     "removal_reason": "",
     "source": "Tobi Lütke (@tobi), chief executive of Shopify, post on X, 21 September 2026",
     "source_url": "https://x.com/tobi/status/2102090718546198790",
     "locator": "Post text; timestamp 21 Sept 2026",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM6",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The incident began on 18 June 2026, when OpenAI's research team used an internal model to research public medicine spending, according to Australia's Prime Minister.",
     "quote": "On June 18, OpenAI's research team used an internal model to conduct internet based research into public medicine spending.",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Prime Minister's answer to the question on how many Australians were affected",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM7",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The agent found a way round repeated blocks on the Medicare statistics portal and accessed public and non-public information in it.",
     "quote": "After encountering repeated blocks … The AI agent found a way around those blocks. … It accessed public and non-public information within the portal",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Prime Minister's answer to the question on how many Australians were affected",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM8",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Services Australia advises that the agent also wrote files to the internal server.",
     "quote": "Services Australia also advises that it engaged, in order to do this, it engaged in writing files as well to the internal server",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Prime Minister's answer to the question on how many Australians were affected",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM9",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI's first notice to the Australian government came on 10 September, as an email to a public mailbox.",
     "quote": "it took until 10 September before there was any notification at all. And the notification was an email sent to just the public mailbox.",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Prime Minister's answer to the question on whether the problem was the delay or the manner of notice",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM10",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "No personal information is believed to have been accessed, and the Prime Minister announced a taskforce to review the incident.",
     "quote": "No personal information is believed to have been accessed at this stage, but investigations are ongoing. … announcing the establishment of a taskforce",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Opening statement, first and second paragraphs",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM11",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Prime Minister said three other systems may have been affected, and that the incident will inform the government's AI standards legislation.",
     "quote": "aware as well of three other systems that may be impacted … insights from this incident will inform the development of our government … AI standards legislation",
     "removal_reason": "",
     "source": "Prime Minister of Australia, Anthony Albanese, \"Press conference - New York\", transcript dated Thursday 24 September 2026",
     "source_url": "https://www.pm.gov.au/media/press-conference-new-york",
     "locator": "Opening statement, second paragraph (legislation) · answer naming AIHW, NSW BOCSAR and the Victorian Department of Health (three systems)",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM12",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic committed about $11.6 billion to Akamai over seven years, with room to expand by up to $9 billion more.",
     "quote": "Anthropic for $11.6 billion of contractual commitment over seven years … expansion of the relationship by up to an additional $9 billion",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Exhibit 99.1 (press release), 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm",
     "locator": "Paragraph 1 (dateline) · second bullet",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM13",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Akamai estimates the capital spending related to the commitment at about $5.5 billion.",
     "quote": "capital expenditures related to today’s $11.6 billion commitment are estimated to be approximately $5.5 billion",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Exhibit 99.1 (press release), 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm",
     "locator": "Paragraph beginning \"Total capital expenditures\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM14",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Akamai expects its 2026 capital spending to rise by about $1.7 billion to buy supply-chain components, including memory, in advance.",
     "quote": "an increase of approximately $1.7 billion in capital expenditures in 2026 to secure and pre-purchase critical supply chain components, including memory",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Exhibit 99.1 (press release), 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm",
     "locator": "Paragraph beginning \"Total capital expenditures\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM15",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Akamai issued Anthropic a warrant for up to about 5% of its common stock, at an exercise price of $111.33 per common share.",
     "quote": "warrant to Anthropic for … up to approximately 5% of Akamai’s common stock outstanding, at an exercise price of $111.33 per share of common stock",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Exhibit 99.1 (press release), 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm",
     "locator": "Paragraph beginning \"As part of the increased strategic alignment\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM16",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The warrant's first tranche, 40% of it, vests on Anthropic's first payment under Project Plan 3.",
     "quote": "the first tranche representing 40% of the Warrant Shares, upon the first payment by Anthropic … to the Company under Project Plan 3",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Item 1.01, filed 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154d8k.htm",
     "locator": "Item 1.01, § Warrant Agreement, paragraph beginning \"The Warrant will vest\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM17",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Each further $3 billion of cloud services Anthropic buys vests about 1% more of Akamai's common stock.",
     "quote": "Each additional $3 billion purchase of cloud services … will result in the vesting of approximately 1% of Akamai’s common stock outstanding.",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Exhibit 99.1 (press release), 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154dex991.htm",
     "locator": "Paragraph beginning \"As part of the increased strategic alignment\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM18",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic may end each project plan on notice of a material outage, subject to conditions.",
     "quote": "Anthropic may terminate each Project Plan upon notice of a material outage, subject to certain conditions.",
     "removal_reason": "",
     "source": "Akamai Technologies, Form 8-K, Item 1.01, filed 24 September 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/1086222/000119312526401048/d288154d8k.htm",
     "locator": "Item 1.01, paragraph beginning \"The Company may terminate the Anthropic MSA\"",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM19",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Project Swap, a book-trading market with 201 Anthropic employees, Claude's ranking of each person's books agreed with the person's own on 61% of pairs, against 50% for random guessing.",
     "quote": "a barter economy with 201 Anthropic employees … Claude’s ordering agreed with theirs 61% of the time (where random guessing would achieve 50%)",
     "removal_reason": "",
     "source": "Anthropic, \"Project Swap: What happens when agents trade for us?\", 24 September 2026",
     "source_url": "https://www.anthropic.com/research/project-swap",
     "locator": "§ Why we did this · § How well Claude guesses preferences",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM20",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM20",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Working from Claude's imprecise rankings caused 85% of the shortfall from the best possible outcome; the trading itself caused the other 15%.",
     "quote": "working from Claude’s imprecise rankings accounts for a majority (85%) of the shortfall … sending agents into a “free-for-all” trading floor accounts for the remaining 15%",
     "removal_reason": "",
     "source": "Anthropic, \"Project Swap: What happens when agents trade for us?\", 24 September 2026",
     "source_url": "https://www.anthropic.com/research/project-swap",
     "locator": "§ The information powering the market limited outcomes",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM21",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Judged on Claude's own rankings, the model an agent ran on made more difference to its outcomes than the instructions it was given.",
     "quote": "the model an agent ran on made more of a difference to its negotiating outcomes than the instructions we gave it",
     "removal_reason": "",
     "source": "Anthropic, \"Project Swap: What happens when agents trade for us?\", 24 September 2026",
     "source_url": "https://www.anthropic.com/research/project-swap",
     "locator": "§ Summary · § On Claude’s own rankings, model choice makes a difference",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM22",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM22",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Participants would let an agent control about 30% of their yearly book budget, against about 40% for a well-read friend.",
     "quote": "The average answer was about 30%. … what share of their budget they would hand over to a well-read friend … The average answer was about 40%.",
     "removal_reason": "",
     "source": "Anthropic, \"Project Swap: What happens when agents trade for us?\", 24 September 2026",
     "source_url": "https://www.anthropic.com/research/project-swap",
     "locator": "§ People generally liked their books, and would give an agent a third of their yearly book budget",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM23",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM23",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5.",
     "quote": "introducing Claude Opus 5.5 … Input and output tokens are $4 and $20 per million, 20% less than Opus 5.",
     "removal_reason": "",
     "source": "Anthropic, \"Introducing Claude Opus 5.5\", 22 September 2026",
     "source_url": "https://www.anthropic.com/claude-opus-5-5",
     "locator": "Opening paragraph · \"Cost and speed\" paragraph · pricing table",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM24",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM24",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic says the new model's costs come out 40% lower than Opus 5's because it costs less per token and uses fewer tokens per task.",
     "quote": "It costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs.",
     "removal_reason": "",
     "source": "Anthropic, \"Introducing Claude Opus 5.5\", 22 September 2026",
     "source_url": "https://www.anthropic.com/claude-opus-5-5",
     "locator": "§ Performance and cost-effectiveness",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM25",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM25",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic attributes the lower price of Opus 5.5 to its lower cost to serve.",
     "quote": "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.",
     "removal_reason": "",
     "source": "Anthropic, \"Introducing Claude Opus 5.5\", 22 September 2026",
     "source_url": "https://www.anthropic.com/claude-opus-5-5",
     "locator": "\"Cost and speed\" paragraph",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM26",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM26",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI priced GPT-6 Sol at $2 per million input tokens and $10 per million output, down from $4 and $20 for GPT-5.6 Sol, and GPT-6 Luna at $0.10 and $0.50, down from $0.20 and $1.20.",
     "quote": "GPT‑5.6 Sol → GPT‑6 Sol $4 → $2 $20 → $10 … GPT‑5.6 Luna → GPT‑6 Luna $0.20 → $0.10 $1.20 → $0.50",
     "removal_reason": "",
     "source": "OpenAI, \"Introducing GPT-6 Sol and Luna\", 22 September 2026",
     "source_url": "https://openai.com/index/introducing-gpt-6-sol-and-luna/",
     "locator": "Table \"GPT-6 API pricing\" (prices per 1 million tokens; columns Input, Output)",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM27",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM27",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI attributes the cut to improvements in caching and inference that lower its cost to serve the models.",
     "quote": "Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on",
     "removal_reason": "",
     "source": "OpenAI, \"Introducing GPT-6 Sol and Luna\", 22 September 2026",
     "source_url": "https://openai.com/index/introducing-gpt-6-sol-and-luna/",
     "locator": "Paragraph above the pricing table",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM28",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM28",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic says Opus 5.5, like Fable 5.1, comes with its watermarking measures to comply with the EU AI Act.",
     "quote": "As with Fable 5.1, Opus 5.5 comes with our watermarking measures to comply with the EU AI Act",
     "removal_reason": "",
     "source": "Anthropic, \"Introducing Claude Opus 5.5\", 22 September 2026",
     "source_url": "https://www.anthropic.com/claude-opus-5-5",
     "locator": "§ Data retention and compliance",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM29",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM29",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic says future Claude models will generate watermarked text, and that it is adding watermarking to older models over the coming months.",
     "quote": "Future Claude models will generate text that contains a watermark. … working to add watermarking for those models … over the coming months",
     "removal_reason": "",
     "source": "Anthropic, \"How Claude's text watermarking works\"",
     "source_url": "https://www.anthropic.com/news/claude-text-watermark",
     "locator": "Opening paragraph · § What about older Claude models?",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM30",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM30",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic's watermark detection API is a private preview, available to eligible organisations as required under EU law.",
     "quote": "We are releasing a detection API in private preview. It is currently available to eligible organizations as required under EU law",
     "removal_reason": "",
     "source": "Anthropic, \"How Claude's text watermarking works\"",
     "source_url": "https://www.anthropic.com/news/claude-text-watermark",
     "locator": "§ How do I check if a piece of text was written by Claude?",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM31",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM31",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Google says every audio clip its Gemini Audio models generate is watermarked with SynthID.",
     "quote": "every audio clip generated by our Gemini Audio models is watermarked with SynthID",
     "removal_reason": "",
     "source": "Google, \"Gemini 3.8 text-to-speech says hello\" (Gemini 3.8 Flash TTS and Flash-Lite TTS), 23 September 2026",
     "source_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/",
     "locator": "§ Build with trust, consent, and transparency",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM32",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM32",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Meta's Muse Realtime Avatar uses Meta Video Seal to embed a durable, invisible watermark in the video it generates.",
     "quote": "Muse Realtime Avatar uses Meta Video Seal to embed a durable, invisible watermark throughout generated video",
     "removal_reason": "",
     "source": "Meta AI Research, \"Bringing Your Muse to Life\", 23 September 2026",
     "source_url": "https://research.meta.ai/blog/bringing-your-muse-to-life",
     "locator": "§ Building Embodied AI Responsibly",
     "date_read": "2026-10-07"
    },
    {
     "id": "AM33",
     "url": "https://durabilitycurve.com/claims/amazon-shut-out-metas-agent/#c-AM33",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Gemini app gives signed-in users a verification tool, including a check of images, video and audio for Google's SynthID watermark.",
     "quote": "verification tool available within the Gemini App for signed-in users … SynthID: Uses invisible watermarks to determine if images, videos, and audio were generated or edited",
     "removal_reason": "",
     "source": "Google, Gemini Apps Help, \"Verify AI-generated images, videos, and audio\"",
     "source_url": "https://support.google.com/gemini/answer/16722517?hl=en",
     "locator": "Opening paragraph · list of checks, item \"SynthID\"",
     "date_read": "2026-10-07"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: One Question for Any AI Model Headline: Can You Use It?",
    "url": "https://durabilitycurve.com/claims/released-to-whom/",
    "essay": {
     "title": "One Question for Any AI Model Headline: Can You Use It?",
     "url": "https://durabilitycurve.com/blog/released-to-whom/",
     "published": "2026-09-25"
    },
    "last_checked": "2026-09-25",
    "format": "current",
    "counts": {
     "claims": 21,
     "by_status": {
      "VERIFIED": 18,
      "EXECUTED": 3
     },
     "verified_at_primary": 18,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/released-to-whom/"
   },
   "rows": [
    {
     "id": "R1",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "At launch Anthropic restricted Mythos 5 to Glasswing partners, with select biology researchers to follow",
     "quote": "\"Claude Mythos 5 is restricted to Glasswing partners (with cyber safeguards lifted) and soon to select biology researchers\"",
     "removal_reason": "",
     "source": "Anthropic, Claude Fable 5 and Claude Mythos 5 (9 Jun 2026)",
     "source_url": "https://www.anthropic.com/news/claude-fable-5-mythos-5",
     "locator": "page-text offset 83.3%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R2",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Mythos 5 is the same underlying model as Fable 5 with safeguards lifted in some areas",
     "quote": "\"It's the same underlying model as Fable 5, but with the safeguards lifted in some areas.\"",
     "removal_reason": "",
     "source": "Anthropic (9 Jun 2026)",
     "source_url": "https://www.anthropic.com/news/claude-fable-5-mythos-5",
     "locator": "page-text offset 6.5%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R3",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic says Mythos 5 has the strongest cybersecurity capabilities of any model",
     "quote": "\"It has the strongest cybersecurity capabilities of any model in the world.\"",
     "removal_reason": "",
     "source": "Anthropic (9 Jun 2026)",
     "source_url": "https://www.anthropic.com/news/claude-fable-5-mythos-5",
     "locator": "page-text offset 7.4%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R4",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic's announcement has an Availability section naming who gets each model (Fable 5 everywhere; Mythos 5 restricted); its subscription-plan terms were staged and later changed, so the help page (R5) carries the current terms",
     "quote": "\"Availability Claude Fable 5 is available everywhere today.\"",
     "removal_reason": "",
     "source": "Anthropic (9 Jun 2026)",
     "source_url": "https://www.anthropic.com/news/claude-fable-5-mythos-5",
     "locator": "page-text offset 83.1%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R5",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On Claude Max, Fable 5 and 5.1 are a standard part of the plan; on Pro they run on pay-as-you-go usage credits (current terms; the separate Fable 5 50% PROMOTION ended 19 Jul 2026 and is not what the article cites)",
     "quote": "\"On Max plans, premium seats on Team plans, and premium seats on legacy seat-based Enterprise plans, both are a standard part of your plan. On Pro plans, standard seats on Team plans, and standard seats on legacy seat-based Enterprise plans, both models run on pay-as-you-go usage credits.\"",
     "removal_reason": "",
     "source": "Claude Help Center, Claude Fable models on your plan (read 2026-09-25)",
     "source_url": "https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan",
     "locator": "page-text offset 74.3%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R6",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On Max, Fable use is capped at up to 50% of weekly usage limits at no extra cost",
     "quote": "\"You can use up to 50% of your weekly usage limits on Fable models at no extra cost.\"",
     "removal_reason": "",
     "source": "Claude Help Center (read 2026-09-25)",
     "source_url": "https://support.claude.com/en/articles/15424964-claude-fable-models-on-your-plan",
     "locator": "page-text offset 77.4%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R7",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI: GPT-6 Astra becomes available to ChatGPT Plus, Pro, Business and Enterprise users (Free not listed)",
     "quote": "\"will become available to all ChatGPT Plus, Pro, Business, and Enterprise users\"",
     "removal_reason": "",
     "source": "OpenAI, GPT-6 Astra (Sept 2026; page undated, carries an \"Update on September 22, 2026\"; page rendered in a browser)",
     "source_url": "https://openai.com/index/gpt-6-astra/",
     "locator": "page-text offset 11.1% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R8",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI: the released Astra refuses more advanced cybersecurity tasks such as proof-of-concept exploits",
     "quote": "\"Astra will refuse to comply with more advanced cybersecurity tasks such as creating proof-of-concept exploits for vulnerabilities.\"",
     "removal_reason": "",
     "source": "OpenAI, GPT-6 Astra (Sept 2026; page undated, carries an \"Update on September 22, 2026\")",
     "source_url": "https://openai.com/index/gpt-6-astra/",
     "locator": "page-text offset 70.8% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R8b",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R8b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI plans to expand access with less restrictive safeguards through Daybreak",
     "quote": "\"Through OpenAI Daybreak⁠, we plan to expand access and roll out less restrictive safeguards in the coming weeks.\"",
     "removal_reason": "",
     "source": "OpenAI, GPT-6 Astra (Sept 2026; page undated, carries an \"Update on September 22, 2026\")",
     "source_url": "https://openai.com/index/gpt-6-astra/",
     "locator": "page-text offset 86.5% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R8c",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R8c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI gives GPT-6 Astra Pro to Pro, Business and Enterprise users (Plus not listed)",
     "quote": "\"Users on the Pro, Business, and Enterprise plans will also get access to GPT‑6 Astra Pro.\"",
     "removal_reason": "",
     "source": "OpenAI, GPT-6 Astra (Sept 2026; page undated, carries an \"Update on September 22, 2026\")",
     "source_url": "https://openai.com/index/gpt-6-astra/",
     "locator": "page-text offset 46.9% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R9",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Google: Gemini 3.8 Flash reaches consumers only on Google AI Pro and Ultra",
     "quote": "\"Consumers: 3.8 Flash is available to Google AI Pro and Ultra subscribers across the Gemini app, AI Mode in Google Search and Gemini in Google Sheets.\"",
     "removal_reason": "",
     "source": "Google, Introducing Gemini 3.8 Flash and 3.8 Flash Cyber (2 Sep 2026)",
     "source_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/",
     "locator": "page-text offset 85.8%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R10",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Google makes Gemini 3.8 Flash Cyber available to trusted defenders through its Fairwind Program",
     "quote": "\"available to trusted defenders through our new Fairwind Program\"",
     "removal_reason": "",
     "source": "Google (2 Sep 2026)",
     "source_url": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/",
     "locator": "page-text offset 34.5%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R11",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "xAI: Grok 4.7 is available in Cursor, Grok Build and the Grok API (the Grok app is not listed)",
     "quote": "\"Grok 4.7 is available today in Cursor and Grok Build. It is also available through the Grok API\"",
     "removal_reason": "",
     "source": "xAI, Introducing Grok 4.7 (21 Sep 2026)",
     "source_url": "https://x.ai/news/grok-4-7",
     "locator": "page-text offset 77.3%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R13",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Solaiman (2023) set out six levels of access to generative AI systems, from fully closed to fully open",
     "quote": "\"six levels of access to generative AI systems: fully closed; gradual or staged access; hosted access; cloud-based or API access; downloadable access; and fully open\"",
     "removal_reason": "",
     "source": "Irene Solaiman, The Gradient of Generative AI Release (arXiv:2302.04844, Feb 2023)",
     "source_url": "https://arxiv.org/abs/2302.04844",
     "locator": "page-text offset 15.3%",
     "date_read": "2026-09-25"
    },
    {
     "id": "R14",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R14",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "On one ChatGPT Plus account (checked 2026-09-25), Chat mode listed GPT-5.6 Sol and GPT-5.5; Work mode showed GPT-6 Astra",
     "quote": "\"GPT-6 Astra Extra High\"",
     "removal_reason": "",
     "source": "First-party observation, the author's own account",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-25"
    },
    {
     "id": "R15",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R15",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "On a Gemini 'Plus' account (checked 2026-09-25), the picker listed 3.5 Flash-Lite, 3.6 Flash, 3.1 Pro and Extended thinking: no 3.8 Flash",
     "quote": "\"3.6 Flash\"",
     "removal_reason": "",
     "source": "First-party observation, the author's own account",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-25"
    },
    {
     "id": "R16",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R16",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Grok's picker (checked 2026-09-25) shows modes (Auto, Fast, Expert, Build, Heavy); hovering shows each mode's model, and every model-naming mode says Grok 4.6",
     "quote": "\"Build apps and sites · Grok 4.6\"",
     "removal_reason": "",
     "source": "First-party observation, the author's own account",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-25"
    },
    {
     "id": "R17",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic: Fable 5.1 and Mythos 5.1 are the same model with different safeguards; Fable 5.1 is generally available, Mythos 5.1 only through trusted access programs",
     "quote": "\"Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is available only through our trusted access programs\"",
     "removal_reason": "",
     "source": "Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1 (1 Sep 2026)",
     "source_url": "https://www.anthropic.com/claude-fable-and-mythos-5-1",
     "locator": "page-text offset 4.0% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R17b",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R17b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic: Mythos 5.1 (cyber safeguards off) has the strongest cyber capabilities of any model Anthropic has released",
     "quote": "\"the strongest cyber capabilities of any model we’ve released\"",
     "removal_reason": "",
     "source": "Anthropic (1 Sep 2026)",
     "source_url": "https://www.anthropic.com/claude-fable-and-mythos-5-1",
     "locator": "page-text offset 63.0% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R18",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI Help: Plus plans include GPT-6 Astra in ChatGPT Work and Codex; GPT-6 Pro (powered by Astra) is for Pro, Business and Enterprise",
     "quote": "\"Plus plans include GPT‑6 Astra in ChatGPT Work and Codex.\"",
     "removal_reason": "",
     "source": "OpenAI Help Center, GPT-5.6 and GPT-6 Pro in ChatGPT (updated ~16 Sep 2026; rendered in a browser)",
     "source_url": "https://help.openai.com/en/articles/20001354-gpt-56-and-gpt-6-pro-in-chatgpt",
     "locator": "page-text offset 72.7% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    },
    {
     "id": "R18b",
     "url": "https://durabilitycurve.com/claims/released-to-whom/#c-R18b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI Help: GPT-6 Pro, powered by GPT-6 Astra, is for Pro, Business and Enterprise plans",
     "quote": "\"GPT‑6 Pro, powered by GPT‑6 Astra, is available in ChatGPT for Pro $100, Pro $200, Business and Enterprise plans\"",
     "removal_reason": "",
     "source": "OpenAI Help Center (rendered in a browser)",
     "source_url": "https://help.openai.com/en/articles/20001354-gpt-56-and-gpt-6-pro-in-chatgpt",
     "locator": "page-text offset 49.2% · page rendered in a browser; availability paragraph saved verbatim",
     "date_read": "2026-09-25"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Use AI to Learn Without Getting Worse at It",
    "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/",
    "essay": {
     "title": "Use AI to Learn Without Getting Worse at It",
     "url": "https://durabilitycurve.com/blog/learn-with-ai-without-getting-worse/",
     "published": "2026-09-24"
    },
    "last_checked": "2026-09-27",
    "format": "current",
    "counts": {
     "claims": 70,
     "by_status": {
      "VERIFIED": 57,
      "EXECUTED": 7,
      "REPORTED": 6
     },
     "verified_at_primary": 57,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/"
   },
   "rows": [
    {
     "id": "B1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a Turkish high-school trial, a plain ChatGPT-style assistant raised practice scores by 48% and a hint-only tutor by 127%",
     "quote": "\"GPT Base and GPT Tutor would increase performance on the assisted practice sessions by 48% and 127%, respectively\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Main Results; Table 1 col (1)",
     "date_read": "2026-09-24"
    },
    {
     "id": "B2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Once the AI was taken away, the plain-assistant group scored 17% lower on the exam than students who never had it",
     "quote": "\"GPT Base diminished the average control student's performance on the unassisted exam by 17%.\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/",
     "locator": "Main Results; Table 1 col (2)",
     "date_read": "2026-09-24"
    },
    {
     "id": "B3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The analysis the authors registered in advance finds a smaller exam harm than the headline (−0.035, CI −0.057 to −0.012), still negative",
     "quote": "\"GPT Base vs. Control ... −0.035 [−0.057, −0.012] 0.003\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025, SI",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12232635/",
     "locator": "SI Appendix A.6, Table A.2",
     "date_read": "2026-09-24"
    },
    {
     "id": "B4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The hint-only tutor removed the exam harm but produced no exam gain",
     "quote": "\"this negative effect is essentially eradicated in the GPT Tutor arm, though we still do not observe a positive effect.\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Main Results; Introduction",
     "date_read": "2026-09-24"
    },
    {
     "id": "B5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The tutor's prompt carried the correct solution and told it not to give the whole solution away",
     "quote": "\"GPT-4 is given a prompt including the solution to each problem (to mitigate hallucinations) as well as instructions to avoid giving away the entire solution\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Footnote †; Experimental Design",
     "date_read": "2026-09-24"
    },
    {
     "id": "B6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Without the answer key, the plain assistant got the practice problems right only 51% of the time",
     "quote": "\"GPT Base gives a correct answer only 51% of the time on average\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "\"GPT Errors vs. Student Performance\"; Fig. 2; SI C.1",
     "date_read": "2026-09-24"
    },
    {
     "id": "B7",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Students used the plain assistant as a crutch, asking for and copying solutions",
     "quote": "\"students often use GPT Base as a \"crutch\" by asking for and copying solutions\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Introduction; Potential Mechanism",
     "date_read": "2026-09-24"
    },
    {
     "id": "B8",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The trial covered nearly 1,000 students",
     "quote": "\"We conducted four 90-min sessions for about fifty 9th, 10th, and 11th-grade classes, comprising nearly 1,000 students.\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Experimental Design",
     "date_read": "2026-09-24"
    },
    {
     "id": "B9",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-B9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The trial measured short-term outcomes only",
     "quote": "\"we focus on short-term outcomes due to limitations imposed by our partner school\"",
     "removal_reason": "",
     "source": "Bastani et al., PNAS 2025",
     "source_url": "https://www.pnas.org/doi/10.1073/pnas.2422633122",
     "locator": "Discussion, final paragraph",
     "date_read": "2026-09-24"
    },
    {
     "id": "K1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a Harvard physics trial, students' median post-test was 4.5 with the AI tutor against 3.5 in an active-learning class",
     "quote": "\"higher median (M) post-score (M = 4.5, N = 142) compared to those in the in-class active learning group (M = 3.5, N = 174)\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Results § Learning gains",
     "date_read": "2026-09-24"
    },
    {
     "id": "K2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Median learning gains with the AI tutor were more than double the class's",
     "quote": "\"in the AI-tutored group were over double those for students in the in-class active learning group\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/",
     "locator": "Results § Learning gains",
     "date_read": "2026-09-24"
    },
    {
     "id": "K3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The regression effect size was 0.63",
     "quote": "\"While the linear regression suggests an effect size of 0.63, this is an underestimation due to ceiling effect\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/",
     "locator": "Results § Linear regression model; Table S1",
     "date_read": "2026-09-24"
    },
    {
     "id": "K4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Harvard tutor's instructions released one step at a time and kept replies brief",
     "quote": "\"Only give away ONE STEP AT A TIME, DO NOT give away the full solution in a single message\" / \"Keep responses BRIEF (a few sentences or less) but helpful.\"",
     "removal_reason": "",
     "source": "Kestin et al., Supplementary Material 1",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Supplement, system prompt",
     "date_read": "2026-09-24"
    },
    {
     "id": "K5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Harvard team loaded each question with its step-by-step worked answer",
     "quote": "\"we enriched our prompts with comprehensive, step-by-step answers\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Discussion § Designing successful student-AI interactions",
     "date_read": "2026-09-24"
    },
    {
     "id": "K6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "194 students took part",
     "quote": "\"Of the 233 enrolled students, 194 were eligible for inclusion in the study.\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Methods § Study population",
     "date_read": "2026-09-24"
    },
    {
     "id": "K9",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Each student learned one lesson with the AI tutor and another in the active-learning class (crossover)",
     "quote": "\"Given that our experiment is a crossover design in which each student experiences both conditions\"",
     "removal_reason": "",
     "source": "Kestin et al., Scientific Reports 2025",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/",
     "locator": "Results (regression controls); Methods § Study design",
     "date_read": "2026-09-24"
    },
    {
     "id": "A1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In an Anthropic trial, 52 developers learned a new Python library, half with an AI assistant",
     "quote": "\"In our main study, 52 participants completed the task, 26 for each of the control and treatment groups.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "§5.2.1",
     "date_read": "2026-09-24"
    },
    {
     "id": "A2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The AI group scored 4.15 points lower on a 27-point quiz, about two grade points",
     "quote": "\"There is a 4.15 point difference between the means ... For a 27-point quiz, this translates into a 17% score difference or 2 grade points.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/html/2601.20245v2",
     "locator": "§5.2.2",
     "date_read": "2026-09-24"
    },
    {
     "id": "A3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Nobody could use AI during the quiz",
     "quote": "\"All participants were not allowed to use AI in the comprehension check.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/html/2601.20245v2",
     "locator": "Figure 4 caption",
     "date_read": "2026-09-24"
    },
    {
     "id": "A4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Using AI did not make them significantly faster",
     "quote": "\"While using AI to complete our coding task did not significantly improve task completion time, the level of skill formation gained ... is significantly reduced.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/html/2601.20245v2",
     "locator": "§1 / §5.2.2",
     "date_read": "2026-09-24"
    },
    {
     "id": "A5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "High-scoring ways of using the AI averaged 65% to 86% on the quiz; low-scoring ways 24% to 39% (post-hoc, small groups)",
     "quote": "\"high-scoring interaction patterns (65%-86% quiz score) vs low-scoring interaction patterns (24%-39% quiz score)\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "§6; Figure 11",
     "date_read": "2026-09-24"
    },
    {
     "id": "A6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Asking the AI conceptual questions was a high-scoring pattern, and the fastest of them",
     "quote": "\"On average, this mode was the fastest among high-scoring patterns and second fastest overall after the AI Delegation mode.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "§6; Figure 11 (Conceptual Inquiry)",
     "date_read": "2026-09-24"
    },
    {
     "id": "N1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a World Bank trial in Nigeria, students had twelve 90-minute sessions guided by teachers",
     "quote": "\"Those assigned to the intervention attended twelve 90-minute sessions in computer labs, engaging in curriculum-aligned activities guided by teachers.\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "§1, p. 2",
     "date_read": "2026-09-24"
    },
    {
     "id": "N2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "English scores rose 0.238 standard deviations, which the authors equate to 1.5 years of ordinary schooling",
     "quote": "\"Our ITT effect of 0.238 standard deviation in English is equivalent to increasing 1.5 years of 'business-as-usual' schooling in Nigeria\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "§4.1, p. 18",
     "date_read": "2026-09-24"
    },
    {
     "id": "N3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the sample week's starting prompt, the chatbot answered the student's question, then set exercises on which it was told to hint after a wrong reply and give the answer only if the reply was still wrong",
     "quote": "\"provide encouraging words and provide a hint\" … \"if my reply is still incorrect\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "Figure 11, PDF p. 51 (Week 7 starting prompt, read from rendered image)",
     "date_read": "2026-09-24"
    },
    {
     "id": "N4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The control group got no programme at all, so extra time and attention are not separated from the AI",
     "quote": "\"the control group, which did not receive any intervention but continued their regular learning in the classroom\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "§2.2, p. 9",
     "date_read": "2026-09-24"
    },
    {
     "id": "N5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors credit the whole package (prompts plus teacher guidance), not the chatbot alone",
     "quote": "\"we interpret that the intervention as a whole -which includes the interaction with the LLM and teacher guidance with specific prompts- is driving the results.\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "§4.3, p. 21",
     "date_read": "2026-09-24"
    },
    {
     "id": "N6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-N6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Nigeria's third-term exam (effect 0.206 SD) came after the intervention on content not limited to it; it ran the day after the last session, so it is not a test of whether the learning lasted",
     "quote": "\"the third term exam score, with an effect size of 0.206 standard deviation (SE = 0.067), although this exam was not limited to the intervention's specific content\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125",
     "source_url": "https://documents1.worldbank.org/curated/en/099548105192529324/pdf/IDU-c09f40d8-9ff8-42dc-b315-591157499be7.pdf",
     "locator": "§4.1; timing from Appendix timeline (pilot sessions end 7/11/24, third-term exam 7/12/24)",
     "date_read": "2026-09-24"
    },
    {
     "id": "O1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI's own trial ran over 300 college students through study sessions of a nominal 40 minutes",
     "quote": "\"we ran a randomized study with over 300 college students preparing for neuroscience and microeconomics exams\" / \"the nominal 40 minute sessions\"",
     "removal_reason": "",
     "source": "OpenAI, \"New tools for understanding AI and learning outcomes\" (2026-03-04)",
     "source_url": "https://openai.com/index/understanding-ai-and-learning-outcomes/",
     "locator": "\"Origins and early research\"; \"Study design\"",
     "date_read": "2026-09-24"
    },
    {
     "id": "O2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Study mode scored roughly 15% higher in microeconomics",
     "quote": "\"roughly a 15% higher score relative\"",
     "removal_reason": "",
     "source": "OpenAI (2026-03-04)",
     "source_url": "https://openai.com/index/understanding-ai-and-learning-outcomes/",
     "locator": "\"Findings\", bullet 2",
     "date_read": "2026-09-24"
    },
    {
     "id": "O3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In neuroscience, study mode was not distinguishable from studying with ordinary online resources",
     "quote": "\"results were not distinguishable from students studying with traditional online resources\"",
     "removal_reason": "",
     "source": "OpenAI (2026-03-04)",
     "source_url": "https://openai.com/index/understanding-ai-and-learning-outcomes/",
     "locator": "\"Findings\", bullet 1",
     "date_read": "2026-09-24"
    },
    {
     "id": "O4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The comparison group used ordinary online resources, not ordinary ChatGPT",
     "quote": "\"a control group studied using traditional online resources such as Google Search and YouTube, with AI generated overview features disabled\"",
     "removal_reason": "",
     "source": "OpenAI (2026-03-04)",
     "source_url": "https://openai.com/index/understanding-ai-and-learning-outcomes/",
     "locator": "\"Study design\", para 1",
     "date_read": "2026-09-24"
    },
    {
     "id": "O5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-O5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI calls the results early; analysis is still underway",
     "quote": "\"While analysis is still underway, early results give us confidence\"",
     "removal_reason": "",
     "source": "OpenAI (2026-03-04)",
     "source_url": "https://openai.com/index/understanding-ai-and-learning-outcomes/",
     "locator": "\"Origins and early research\"",
     "date_read": "2026-09-24"
    },
    {
     "id": "G1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Google's own trial of Gemini Guided Learning in Sierra Leone raised maths scores 0.258 standard deviations (CI 0.027 to 0.488)",
     "quote": "\"yielding a gain of +0.258 standard deviations across the Guided Learning classrooms (intent to treat; 95% confidence interval [0.027, 0.488]\"",
     "removal_reason": "",
     "source": "LearnLM Team, Google & Fab AI, tech report (May 2026)",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "p. 3; Table C.4",
     "date_read": "2026-09-24"
    },
    {
     "id": "G2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "It enrolled 1,763 students in 48 grade 7 and 8 classrooms",
     "quote": "\"The trial enrolled 𝑁= 1763 students aged 13 or older in 48 grades 7 and 8 classrooms.\"",
     "removal_reason": "",
     "source": "LearnLM Team tech report",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "p. 2",
     "date_read": "2026-09-24"
    },
    {
     "id": "G3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Students shared devices in pairs, in lessons teachers ran",
     "quote": "\"Students accessed the Gemini app on tablets or desktop computers, sharing at a 2:1 student-to-device ratio.\"",
     "removal_reason": "",
     "source": "LearnLM Team tech report",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "p. 2",
     "date_read": "2026-09-24"
    },
    {
     "id": "G4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The gain came from grade 8; the grade 7 estimate was slightly negative (−0.078; grade 8 interaction 0.429)",
     "quote": "\"Arm: Treatment –0.078* (0.037)\"; \"Arm: Treatment × Grade: Grade 8 0.429** (0.142)\"",
     "removal_reason": "",
     "source": "LearnLM Team tech report",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "Table C.14, p. 28",
     "date_read": "2026-09-24"
    },
    {
     "id": "G5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Gemini gave a direct solution in 2.1% of its messages (Gemini-classified)",
     "quote": "\"far more frequently than providing direct solutions (2.1%)\"",
     "removal_reason": "",
     "source": "LearnLM Team tech report",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "p. 4",
     "date_read": "2026-09-24"
    },
    {
     "id": "G6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-G6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The report has not been peer reviewed",
     "quote": "\"Please note works submitted as a preprint have not undergone a peer review process.\"",
     "removal_reason": "",
     "source": "LearnLM Team tech report",
     "source_url": "https://storage.googleapis.com/deepmind-media/LearnLM/learnLM_sierraleone_may26.pdf",
     "locator": "p. 8",
     "date_read": "2026-09-24"
    },
    {
     "id": "D1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "OpenAI's own help page says Study mode may still give a direct answer",
     "quote": "\"there may be times when it gives a direct answer\"",
     "removal_reason": "",
     "source": "OpenAI Help Center, \"Using study mode in ChatGPT\"",
     "source_url": "https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Study mode is not available inside ChatGPT Projects",
     "quote": "\"The Study option is not available in Temporary Chats, GPTs, or Projects.\"",
     "removal_reason": "",
     "source": "OpenAI Help Center",
     "source_url": "https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A ChatGPT scheduled task made inside a project cannot read that project's files",
     "quote": "\"If you create a task in a project, it cannot access uploaded files or files stored in that project.\"",
     "removal_reason": "",
     "source": "OpenAI Help Center, Tasks",
     "source_url": "https://help.openai.com/en/articles/10291617",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude's free plan allows up to five Projects",
     "quote": "\"Free users can create a maximum of five projects.\"",
     "removal_reason": "",
     "source": "Claude Help Center",
     "source_url": "https://support.claude.com/en/articles/9519177",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Gemini Guided Learning sits under Add files, then More tools",
     "quote": "\"In the text box, click Add Files.\" / \"At the bottom, click More tools\" … \"Guided Learning\"",
     "removal_reason": "",
     "source": "Google Gemini Help",
     "source_url": "https://support.google.com/gemini/answer/16448384",
     "locator": "\"Use Guided Learning\", steps 2–3",
     "date_read": "2026-09-27"
    },
    {
     "id": "X1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Study appears as a composer mode in the ChatGPT web app (Plus)",
     "quote": "The author's screenshot (not published): composer shows \"Study\" mode chip",
     "removal_reason": "",
     "source": "First-party: ChatGPT web app (Plus), the author's run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "X2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In 8 runs of the final Level 2 prompt (7 across Claude, Gemini and GPT, 1 in the ChatGPT app), the tutor withheld the answer on a direct demand and, after the second wrong try, gave the answer (some models per step) and named the missed step; once (GPT, code) it described the fix in words while withholding the code",
     "quote": "See run record",
     "removal_reason": "",
     "source": "First-party: python3 run.py, run_gemini.py, run_codex.py + the author's ChatGPT run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "X3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X3",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In a simulated project (instructions plus file given to the model directly), the Level 3 setup stuck to the file, said when the file did not cover a question, and scored a no-feedback quiz correctly on all three model families; the Claude and Gemini apps themselves were not tested",
     "quote": "See run record",
     "removal_reason": "",
     "source": "First-party: python3 l3.py, l3x.py",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "X4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X4",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In the ChatGPT app, our first Level 3 instructions answered a \"just tell me\" question at once from general knowledge; the revised instructions held it back and answered from the file",
     "quote": "See run record",
     "removal_reason": "",
     "source": "First-party: ChatGPT web (Plus), Project \"Cold Check Test\", the author's run (v2) + Claude via Claude in Chrome (v3)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "X5",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X5",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In the ChatGPT app, the cold-check prompt held all feedback to the end and scored a 3-of-6 quiz correctly against the file",
     "quote": "See run record",
     "removal_reason": "",
     "source": "First-party: ChatGPT web (Plus), Project \"Cold Check Test\", driven by Claude via Claude in Chrome",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "D6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Claude.ai, the old Styles menu is gone; Anthropic now offers an official \"learn\" skill (Customize → Skills → Discover → learn → Add) whose stated goal is to help the learner answer it themselves, and which does not trigger on coding tasks or factual lookups",
     "quote": "\"The goal is not to answer the learner's question but to help them be able to answer it themselves\" / \"Don't trigger for: Tasks: coding, writing, calculation, translation, factual lookup\"",
     "removal_reason": "",
     "source": "Anthropic \"learn\" skill (updated Sep 14), SKILL.md + Overview, read in the live Claude.ai app",
     "source_url": "https://claude.ai/new#customize/skills/id/learn",
     "locator": "Skill Overview + Contents › SKILL.md \"Learning Mode\"",
     "date_read": "2026-09-24"
    },
    {
     "id": "A7",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A7",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "They were Python programmers recruited through a crowd-work platform; 29 of the 52 had seven or more years of coding experience",
     "quote": "Paper Table 1 (experience bands; 7+ years = 29 of 52) / \"the small group size of the 1-3 year participant group (n=4)\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "Table 1; §5.2.2",
     "date_read": "2026-09-24"
    },
    {
     "id": "S2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-S2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Only two of the seven studies went through journal peer review (PNAS and Scientific Reports); the others are two preprints, a working paper, a vendor tech report and a vendor blog post",
     "quote": "PNAS 122(26); Scientific Reports 15, 17458; arXiv 2601.20245 and arXiv 2409.09047 (preprints); World Bank Policy Research Working Paper 11125; Google tech report \"have not undergone a peer review process\"; OpenAI \"While analysis is still underway\"",
     "removal_reason": "",
     "source": "The seven studies' own publication records",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "D7",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D7",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Microsoft Copilot has a \"Study and learn\" mode, chosen from the Quick response menu under the prompt",
     "quote": "\"Press (or tap) 'Quick response' under your prompt and select an appropriate mode\"",
     "removal_reason": "",
     "source": "Microsoft Support, \"Conversation modes in Microsoft Copilot\"",
     "source_url": "https://support.microsoft.com/en-us/microsoft-copilot/conversation-modes-in-microsoft-copilot",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D8",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D8",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Claude Code has a Learning output style, switched on with /output-style learning",
     "quote": "\"/output-style learning\" (Learning style: Claude \"stops and waits\" at TODO(human) markers)",
     "removal_reason": "",
     "source": "Claude Code docs, Output styles",
     "source_url": "https://code.claude.com/docs/en/output-styles",
     "locator": "Built-in styles section",
     "date_read": "2026-09-24"
    },
    {
     "id": "K7",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Harvard tutor's instructions allowed it to give the answer if the student demanded it",
     "quote": "\"DO NOT not tell them the answer UNLESS they demand you to give them the answer\"",
     "removal_reason": "",
     "source": "Kestin et al., Supplementary Material 1",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Supplement, system prompt",
     "date_read": "2026-09-24"
    },
    {
     "id": "K8",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-K8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Harvard tutor's instructions asked students to try first",
     "quote": "\"encourage them to give it a try first\"",
     "removal_reason": "",
     "source": "Kestin et al., Supplementary Material 1",
     "source_url": "https://www.nature.com/articles/s41598-025-97652-6",
     "locator": "Supplement, system prompt",
     "date_read": "2026-09-24"
    },
    {
     "id": "A8",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The highest-scoring pattern (2 people) had the AI generate code and then asked it questions to understand it, averaging 86%; asking conceptual questions averaged 65%; the lowest patterns, which delegated or relied on it, averaged 24% to 39%",
     "quote": "\"high-scoring interaction patterns (65%-86% quiz score) vs low-scoring interaction patterns (24%-39% quiz score)\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "§1.1; §6; Figure 11 (Generation-Then-Comprehension n=2 86%; Conceptual Inquiry n=7 65%)",
     "date_read": "2026-09-24"
    },
    {
     "id": "A10",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The low-scoring patterns were delegation, progressive reliance and iterative AI debugging (relying on the AI to debug or verify their code)",
     "quote": "\"Iterative AI Debugging (n=4): Participants in this group relied on AI to debug or verify their code.\"",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245v2",
     "source_url": "https://arxiv.org/html/2601.20245v2",
     "locator": "§6, interaction patterns",
     "date_read": "2026-09-24"
    },
    {
     "id": "L1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In pre-registered lab experiments on learning to code, giving students an AI chatbot had no overall effect on learning",
     "quote": "\"we find no effect of LLMs on overall learning outcomes\"",
     "removal_reason": "",
     "source": "Lehmann, Cornelius & Sting, arXiv 2409.09047v2",
     "source_url": "https://arxiv.org/abs/2409.09047",
     "locator": "Abstract; Tables 5, 7",
     "date_read": "2026-09-24"
    },
    {
     "id": "L2",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Students who used it to substitute for their own work (e.g. generating solutions to exercises) covered more topics but understood less; students who used it to complement their work (e.g. asking for explanations) understood more (exploratory, not randomised)",
     "quote": "\"Students who substitute some of their learning activities with LLMs (e.g., by generating solutions to exercises) increase the volume of topics\" / \"Students who complement their learning activities with LLMs (e.g., by asking for explanations) do not increase topic volume but do increase their understanding.\"",
     "removal_reason": "",
     "source": "Lehmann et al., arXiv 2409.09047v2",
     "source_url": "https://arxiv.org/abs/2409.09047",
     "locator": "Abstract; §9",
     "date_read": "2026-09-24"
    },
    {
     "id": "L3",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "42% of messages asking for a solution were sent without a single attempt",
     "quote": "\"42% of messages asking for a solution were sent without a single attempt\"",
     "removal_reason": "",
     "source": "Lehmann et al., arXiv 2409.09047v2",
     "source_url": "https://arxiv.org/abs/2409.09047",
     "locator": "§9.1 (Study 3)",
     "date_read": "2026-09-24"
    },
    {
     "id": "L4",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-L4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The chatbot raised how much students felt they had learned by more than their actual learning explains",
     "quote": "\"LLMs increase perceived learning by more than can be explained by actual differences in learning\"",
     "removal_reason": "",
     "source": "Lehmann et al., arXiv 2409.09047v2",
     "source_url": "https://arxiv.org/abs/2409.09047",
     "locator": "§9.3",
     "date_read": "2026-09-24"
    },
    {
     "id": "D9",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude Skills need code execution switched on (Settings > Capabilities > Code execution and file creation on Free, Pro and Max)",
     "quote": "\"This feature requires code execution to be enabled.\"",
     "removal_reason": "",
     "source": "Claude Help Center, \"Use Skills in Claude\"",
     "source_url": "https://support.claude.com/en/articles/12512180-use-skills-in-claude",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "D10",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D10",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "ChatGPT's Study mode is chosen by typing @study or from the + menu, and is on every plan",
     "quote": "\"Available across ChatGPT plans globally on web, iOS, and Android\"",
     "removal_reason": "",
     "source": "OpenAI Help Center, \"Using study mode in ChatGPT\"",
     "source_url": "https://help.openai.com/en/articles/11780217-using-study-mode-in-chatgpt",
     "locator": "Article body",
     "date_read": "2026-09-24"
    },
    {
     "id": "X6",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-X6",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The cold check with a topic line kept to the named topics and scored correctly in the ChatGPT app (3 of 6, the planted misses exactly), and kept to topic on 17 of 18 questions across Claude, Gemini and GPT",
     "quote": "See run record",
     "removal_reason": "",
     "source": "First-party: ChatGPT web (Plus) via Claude in Chrome + l3_cc2.py, l3x_cc2.py",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-24"
    },
    {
     "id": "P1",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-P1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Tutor prompts built to make students work before giving answers were published by Ethan and Lilach Mollick in 2023, and the Nigeria programme borrowed from them",
     "quote": "\"Some of the prompt structures were derived from Mollick and Mollick (2023a)\"",
     "removal_reason": "",
     "source": "De Simone et al., World Bank PRWP 11125, fn 3; Mollick & Mollick, \"Assigning AI\" (SSRN 4475995)",
     "source_url": "https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4475995",
     "locator": "§2.1, pp. 7 to 8 (main text)",
     "date_read": "2026-09-24"
    },
    {
     "id": "A9",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-A9",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The no-AI group averaged about 65% on the quiz",
     "quote": "Figure 6 plots about 65.4% for the control group (read from marker positions); the Anthropic blog gives 67%",
     "removal_reason": "",
     "source": "Shen & Tamkin, arXiv 2601.20245, Figure 6; Anthropic blog",
     "source_url": "https://arxiv.org/abs/2601.20245",
     "locator": "Figure 6",
     "date_read": "2026-09-24"
    },
    {
     "id": "D11",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "ChatGPT project: New project in the sidebar; instructions via the project's settings",
     "quote": "\"Select New project in the sidebar.\" / \"Select the more options menu (•••), then select Project settings to add instructions for the project.\"",
     "removal_reason": "",
     "source": "OpenAI Help Center, Projects in ChatGPT",
     "source_url": "https://help.openai.com/en/articles/10169521",
     "locator": "\"Create a project\"; \"Add project instructions\"",
     "date_read": "2026-09-27"
    },
    {
     "id": "D12",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude project: Projects, then New Project, with instructions and knowledge files",
     "quote": "\"click “Projects,”\" … \"Click \"+ New Project\" in the upper right corner.\" / \"Anything you upload to this space will be used across all of your chats within that project.\" / \"Click on \"Set project instructions.\"\"",
     "removal_reason": "",
     "source": "Claude Help Center, projects article",
     "source_url": "https://support.claude.com/en/articles/9519177",
     "locator": "\"How to create a project\"; \"Add content to project knowledge\"; \"Add project instructions\"",
     "date_read": "2026-09-27"
    },
    {
     "id": "D13",
     "url": "https://durabilitycurve.com/claims/learn-with-ai-without-getting-worse/#c-D13",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Gemini: a Gem (Gems, then New Gem) with files added under Knowledge",
     "quote": "Help-page path: Gems → New Gem → Knowledge → Add files",
     "removal_reason": "",
     "source": "Google Gemini Help",
     "source_url": "https://support.google.com/gemini/answer/15146780",
     "locator": "Article body",
     "date_read": "2026-09-24"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Your Health Has a Dashboard. Your Relationships Don't.",
    "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/",
    "essay": {
     "title": "Your Health Has a Dashboard. Your Relationships Don't.",
     "url": "https://durabilitycurve.com/blog/your-health-has-a-dashboard/",
     "published": "2026-09-23"
    },
    "last_checked": "2026-09-23",
    "format": "current",
    "counts": {
     "claims": 10,
     "by_status": {
      "VERIFIED": 10
     },
     "verified_at_primary": 10,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/your-health-has-a-dashboard/"
   },
   "rows": [
    {
     "id": "C1",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The meta-analysis pooled 148 studies and 308,849 participants",
     "quote": "\"Across 148 studies (308,849 participants), the random effects weighted average effect size was OR = 1.50\"",
     "removal_reason": "",
     "source": "Holt-Lunstad, Smith & Layton 2010, PLoS Medicine 7(7):e1000316",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Abstract (Results)",
     "date_read": "2026-09-21"
    },
    {
     "id": "C2",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Participants were followed for an average of seven and a half years",
     "quote": "\"Participants were followed for an average of 7.5 years (SD = 7.1, range = 3 months to 58 years)\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Abstract (Methods)",
     "date_read": "2026-09-21"
    },
    {
     "id": "C3",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Stronger social relationships → a 50% increased likelihood of survival, OR 1.50 (95% CI 1.42–1.59)",
     "quote": "\"OR = 1.50 (95% CI 1.42 to 1.59)\" and \"a 50% increased likelihood of survival for participants with stronger social relationships\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Abstract (Results / Conclusions)",
     "date_read": "2026-09-21"
    },
    {
     "id": "C4",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Complex measures of social integration carried the largest pooled association, OR 1.91 (95% CI 1.63–2.23)",
     "quote": "\"OR = 1.91; 95% CI 1.63 to 2.23\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Results — moderator analysis by type of assessment (complex measures of social integration)",
     "date_read": "2026-09-21"
    },
    {
     "id": "C5",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The meta-analysis reported substantial between-study heterogeneity, I²=81%",
     "quote": "\"I2 = 81% [95% CI = 78% to 84%]\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Results (overall heterogeneity)",
     "date_read": "2026-09-21"
    },
    {
     "id": "C6",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The 2010 paper framed the effect-size magnitude as comparable to quitting smoking",
     "quote": "\"The magnitude of this effect is comparable with quitting smoking and it exceeds many well-known risk factors for mortality\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Discussion",
     "date_read": "2026-09-21"
    },
    {
     "id": "C7",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors issued a 2024 author note correcting the Figure 6 cross-risk comparisons; the relationship result itself stood",
     "quote": "\"our original conclusion about Figure 6 in the text of the 2010 publication remains consistent: the magnitude of the effect of social relationships is on par with other widely recognized risk factors for mortality\" (author note, posted 30 April 2024; error was inconsistent risk-ratio/odds-ratio/hazard-ratio conversion across the comparison meta-analyses)",
     "removal_reason": "",
     "source": "Holt-Lunstad et al., Author Note: Figure 6, PLOS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article/comment?id=10.1371/annotation/e56c92ac-9c2f-4d4e-9e96-b092b8b967b7",
     "locator": "Author note / annotation, 30 April 2024",
     "date_read": "2026-09-21"
    },
    {
     "id": "C8",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors noted that statistical controls used by many included studies rule out a number of potential confounders; the association held after adjustment, though controlled effect sizes ran lower in magnitude than raw (backs §4 \"adjusted for many of these and the association held, but adjustment is not proof\")",
     "quote": "\"Statistical controls (e.g., age, sex, physical condition, etc.) employed by many of the studies rule out a number of potentially confounding variables that might account for the association between social relationships and mortality\" (Discussion); \"effect sizes based on data controlling for other variables were lower in magnitude than those based on raw data\" (Results)",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC2910600/",
     "locator": "Discussion (statistical controls) + Results (controlled vs raw)",
     "date_read": "2026-09-22"
    },
    {
     "id": "C9",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A within-person daily study found people reported greater happiness and belonging on days with more weak-tie / acquaintance contact; weak ties relate to social and emotional well-being. OUTCOME = well-being, NOT survival (used in §5 to support breadth, scoped as a different outcome from the mortality meta)",
     "quote": "\"In Study 1, students experienced greater happiness and greater feelings of belonging on days when they interacted with more classmates than usual... we again found that weak ties are related to social and emotional well-being. The current results highlight the power of weak ties, suggesting that even social interactions with the more peripheral members of our social networks contribute to our well-being.\"",
     "removal_reason": "",
     "source": "Sandstrom & Dunn 2014, Personality and Social Psychology Bulletin 40(7):910–922",
     "source_url": "https://repository.essex.ac.uk/15300/",
     "locator": "Abstract",
     "date_read": "2026-09-22"
    },
    {
     "id": "C10",
     "url": "https://durabilitycurve.com/claims/your-health-has-a-dashboard/#c-C10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The lowest category in the moderator analysis was a binary living-alone measure, OR 1.19 (footnote 3)",
     "quote": "\"the association was strongest for complex measures of social integration (OR = 1.91; 95% CI 1.63 to 2.23) and lowest for binary indicators of residential status (living alone versus with others) (OR = 1.19\"",
     "removal_reason": "",
     "source": "Holt-Lunstad et al. 2010, PLoS Medicine",
     "source_url": "https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1000316",
     "locator": "Abstract (Results)",
     "date_read": "2026-09-23"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: You Can Win the Wrong Game for Years",
    "url": "https://durabilitycurve.com/claims/the-wrong-game/",
    "essay": {
     "title": "You Can Win the Wrong Game for Years",
     "url": "https://durabilitycurve.com/blog/the-wrong-game/",
     "published": "2026-09-14"
    },
    "last_checked": "2026-09-27",
    "format": "current",
    "counts": {
     "claims": 17,
     "by_status": {
      "VERIFIED": 11,
      "REPORTED": 6
     },
     "verified_at_primary": 11,
     "removed": 1
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-wrong-game/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "By Palmer's count there were twelve trials of scientists over their science, Galileo's among them, and only one ended in execution (Giordano Bruno)",
     "quote": "\"There are 12 total trials of scientists about science. Galileo is one. Giordano Bruno is one. Giordano Bruno is the only one executed. Of those 12 trials, only three were convicted.\"",
     "removal_reason": "",
     "source": "Ada Palmer, spoken, Dwarkesh Patel podcast",
     "source_url": "https://dwarkesh.com/p/ada-palmer",
     "locator": "Transcript, chapter \"The Inquisition accidentally invented peer review\" (01:41:21), her reply to Dwarkesh on executions",
     "date_read": "2026-09-13"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-2",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "2026-09-13: the \"Lutherans and Calvinists … only executed one person for doing science\" sentence is the episode EDITOR BLURB, not Palmer's speech (the prior row-1 locator was wrong). The peer-review point IS also in her speech (\"they effectively invented peer review\", crediting Macuglia). Neither blurb wording is used in the body.",
     "quote": "\"The focus of the Inquisition is really misunderstood - it was obsessed with catching dangerous new heretics like Lutherans and Calvinists - it only executed one person for doing science.\"",
     "removal_reason": "",
     "source": "Dwarkesh × Ada Palmer — EDITOR SUMMARY",
     "source_url": "https://dwarkesh.com/p/ada-palmer",
     "locator": "Episode summary blurb, above the chapter list",
     "date_read": "2026-09-13"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-3",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Inquisitors cared about the fine print of Protestant theology; a 1545 index of banned books promised to print arch-heretics' names in capitals and gave them to Calvin, Zwingli, Luther and Melanchthon; Machiavelli was not in capitals (\"not important enough from their position\"). Attributed in body to Palmer (\"She points to…\").",
     "quote": "",
     "removal_reason": "A 1545 index existed (Paris theology faculty) but was not the Inquisition's, and the capitals/Machiavelli detail could not be confirmed at a second source. The essay now uses her verbatim Lucretius letters + \"all of these fine minutiae of Protestantism\" instead (same transcript passage).",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-09-13"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Investors think buying the stocks is a bet on the technology; Dalio says it is not",
     "quote": "\"What a lot of people don't realize in bubbles is that through all technologies, they think that they are betting on the technology when they buy the stocks in the companies. That's not true.\"",
     "removal_reason": "",
     "source": "Ray Dalio on the All-In Podcast (reported)",
     "source_url": "https://finance.yahoo.com/markets/stocks/articles/ray-dalio-says-ai-investors-121700871.html",
     "locator": "Article body",
     "date_read": "2026-09-12"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Dalio's mechanism: technology success and company survival are different behaviours; most early companies do not survive",
     "quote": "\"There's a giant difference between the behavior of companies and the behavior of the technologies. The norm is … a lot of companies won't survive in the start. Very small percentage.\"",
     "removal_reason": "",
     "source": "Ray Dalio on the All-In Podcast (reported)",
     "source_url": "https://finance.yahoo.com/markets/stocks/articles/ray-dalio-says-ai-investors-121700871.html",
     "locator": "Article body",
     "date_read": "2026-09-12"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-6",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The chemical surge of an emotion clears in about 90 seconds; what sustains it past that is re-triggering it by narrating — so the game is attention direction, not fighting the feeling. HER SUBJECTIVE ACCOUNT, not controlled research",
     "quote": "\"it takes less than 90 seconds\" for the chemical rush of an emotion to surge and flush; \"anything you continue to feel is your own choosing\" (paraphrase of her framing; not controlled research per multiple summaries)",
     "removal_reason": "",
     "source": "Jill Bolte Taylor, My Stroke of Insight (via Steven Bartlett podcast; shortform summary)",
     "source_url": "https://shortform.com/blog/jill-bolte-taylors-90-second-rule",
     "locator": "Book summary; her subjective account",
     "date_read": "2026-09-12"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-7",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Next-token prediction was widely called the wrong game for intelligence, and scaling that crude objective produced strong capability — so a confident \"wrong game\" verdict can itself be the targeting error when the crude game is a sufficient statistic",
     "quote": "(interpretive/argument claim; our archive's amended Law V case, not an external figure)",
     "removal_reason": "",
     "source": "The Durability Curve's own framework",
     "source_url": "",
     "locator": "worldview.md § Law V, facet (b)",
     "date_read": "2026-09-12"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "SECOND reversal case (non-AI): a crude statistic the sophisticated baseball market underpriced turned out to track winning; the A's exploited the inefficiency and the market then corrected — the \"look at the whole player\" instinct was the misread",
     "quote": "\"These methods support Lewis's argument that certain baseball skills were valued inefficiently in the early part of this period, and that this inefficiency was profitably exploited by managers with the ability to generate and interpret statistical knowledge. Consistent with Lewis's story and economic reasoning, as knowledge of the inefficiency became increasingly dispersed across baseball teams the market corrected the original mispricing.\"",
     "removal_reason": "",
     "source": "Hakes & Sauer, \"An Economic Evaluation of the Moneyball Hypothesis\", Journal of Economic Perspectives 20(3):173–186, 2006",
     "source_url": "https://aeaweb.org/articles?id=10.1257/jep.20.3.173",
     "locator": "Abstract",
     "date_read": "2026-09-12"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-9",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The specific undervalued skill was on-base percentage / drawing walks (the paper's body finds OBP ~2× as important as slugging for runs and wins); traditional scouts derided the statistical approach in favour of gut, five-tool, whole-player judgement",
     "quote": "(OBP-vs-SLG magnitude reported from the paper's body/AEA summary, not read in the abstract; the scouts-vs-stats clash is the documented Moneyball account)",
     "removal_reason": "",
     "source": "Hakes & Sauer (body) + Michael Lewis, Moneyball (2003)",
     "source_url": "https://aeaweb.org/articles?id=10.1257/jep.20.3.173",
     "locator": "Body findings + Lewis",
     "date_read": "2026-09-12"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-10",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Next-token prediction looked unimpressive as a route to \"understanding\" for years and capability rose with scale over time (the GPT-2 → GPT-3/4 arc), so for a long stretch the sceptics' evidence looked like it supported them.",
     "quote": "",
     "removal_reason": "",
     "source": "ML history — scaling-laws arc (Kaplan et al. 2020; GPT-2 2019 → GPT-4)",
     "source_url": "",
     "locator": "historical arc, not a cited number",
     "date_read": "2026-09-12"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Palmer's general conclusion: censors are always wrong, from our perspective, about what they should be worried about",
     "quote": "\"whatever they're looking at, they're always wrong, from our perspective, about what they should be worried about censoring\"",
     "removal_reason": "",
     "source": "Ada Palmer, spoken",
     "source_url": "https://dwarkesh.com/p/ada-palmer",
     "locator": "Transcript, same chapter (\"One fun game when I study the history of censorship…\")",
     "date_read": "2026-09-13"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Palmer's next NON-FICTION book is a history of censorship (body: \"whose next non-fiction book is a history of censorship\"; she also writes fiction)",
     "quote": "\"my next non-fiction book is gonna be a book on the history of censorship\"",
     "removal_reason": "",
     "source": "Ada Palmer, spoken",
     "source_url": "https://dwarkesh.com/p/ada-palmer",
     "locator": "Transcript, same passage",
     "date_read": "2026-09-13"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The 2002 Oakland A's won as many games as the 2002 Yankees (103 each)",
     "quote": "\"a record of 103–59\" (A's); Yankees 103–58 in the standings table",
     "removal_reason": "",
     "source": "Wikipedia, 2002 Oakland Athletics season",
     "source_url": "https://en.wikipedia.org/wiki/2002_Oakland_Athletics_season",
     "locator": "Infobox + standings table",
     "date_read": "2026-09-13"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Yankees' 2002 payroll was about three times the A's (A's $39,679,746 vs Yankees $125,928,583 on Opening Day)",
     "quote": "\"Yankees 103-58 $125,928,583 $133,429,575\" · \"Athletics 103-59 $39,679,746 $41,942,665\" (Opening Day and August 31 payroll columns)",
     "removal_reason": "",
     "source": "Doug Pappas, SABR Business of Baseball, 2002 payroll table",
     "source_url": "https://roadsidephotos.sabr.org/baseball/02-4pay.htm",
     "locator": "payroll table, Opening Day and August 31 columns",
     "date_read": "2026-09-27"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Prior art credited in body: Annie Duke's kill criteria pair a state and a date",
     "quote": "\"Kill criteria, generally, include both states and dates\"",
     "removal_reason": "",
     "source": "Annie Duke, Quit (2022), via Behavioral Scientist excerpt",
     "source_url": "https://behavioralscientist.org/annie-duke-quit-mental-models-to-help-you-cut-your-losses/",
     "locator": "Article body, kill-criteria section",
     "date_read": "2026-09-13"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The next-token sceptics argued the approach lacked understanding/meaning (dated examples)",
     "quote": "Bender & Koller: \"we argue that a system trained only on form has a priori no way to learn meaning\" · Marcus: \"prediction does not equal understanding\"",
     "removal_reason": "",
     "source": "Bender & Koller, \"Climbing towards NLU\", ACL 2020; Marcus, \"GPT-2 and the Nature of Intelligence\", The Gradient, 25 Jan 2020",
     "source_url": "https://aclanthology.org/2020.acl-main.463/",
     "locator": "Bender & Koller, abstract; Marcus, essay body",
     "date_read": "2026-09-27"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-17",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Many companies that went public on the internet wave in 1999 did not survive the crash (body: \"a great many\", no figure)",
     "quote": "Webvan (IPO Nov 1999, bankrupt Jul 2001) and eToys (IPO May 1999, Chapter 11 Feb 2001) among CNBC's failed dot-com IPOs; ~4,800 of 7–10k late-90s online ventures sold or gone by mid-2003",
     "removal_reason": "",
     "source": "CNBC, \"Failed IPOs of the Dot-Com Bubble\"",
     "source_url": "https://cnbc.com/2012/05/17/Failed-IPOs-of-the-Dot-Com-Bubble.html",
     "locator": "Slideshow entries",
     "date_read": "2026-09-13"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/the-wrong-game/#c-18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Palmer quotes inquisitors' letters: no need to censor Lucretius (only learned people can read it); what needs censoring is \"all of these fine minutiae of Protestantism\"",
     "quote": "\"We don't need to bother censoring Lucretius. Only learned people can read it\" … \"What we need to worry about censoring is all of these fine minutiae of Protestantism.\"",
     "removal_reason": "",
     "source": "Ada Palmer, spoken (paraphrasing inquisitors' letters)",
     "source_url": "https://www.dwarkesh.com/p/ada-palmer",
     "locator": "Transcript, censorship passage (chapter at 01:41:21)",
     "date_read": "2026-09-13"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Planes That Didn't Come Back",
    "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/",
    "essay": {
     "title": "The Planes That Didn't Come Back",
     "url": "https://durabilitycurve.com/blog/the-planes-that-didnt-come-back/",
     "published": "2026-09-09"
    },
    "last_checked": "2026-10-13",
    "format": "current",
    "counts": {
     "claims": 7,
     "by_status": {
      "VERIFIED": 7
     },
     "verified_at_primary": 7,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/"
   },
   "rows": [
    {
     "id": "W1",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Wald's work on aircraft vulnerability is a series of eight memoranda, originally published in 1943 by the Statistical Research Group at Columbia University for the National Defense Research Committee. Scope: these are the words of the abstract that the Center for Naval Analyses wrote for its 1980 reprint, not Wald's own.",
     "quote": "a series of eight memoranda originally published by the Statistical Research Group at Columbia University for the National Defense Research Committee in 1943",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors (eight memoranda, 1943), reprinted by the Center for Naval Analyses as CRC 432, July 1980; abstract on the reprint's Report Documentation Page",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Report Documentation Page, block 20 (Abstract), which follows printed p. 89 and is outside the body's page numbering (PDF p. 99). Block 18 on the same page lists the eight memo numbers.",
     "date_read": "2026-10-13"
    },
    {
     "id": "W2",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Wald's method starts from two kinds of count: the total number of planes that took part in combat and, for each number of hits, the number of planes that took exactly that many hits and returned. Scope: the memoranda define their data in these terms and do not say who examined the aircraft or how the hits were recorded.",
     "quote": "The total number N of planes participating in combat … planes that received exactly i hits but have not been downed, i.e., have returned from combat",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part I, An Equation Satisfied by the Probabilities That a Plane Will Be Downed by i Hits (SRG memo 85), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part I, Introduction, printed p. 1 (PDF p. 10)",
     "date_read": "2026-10-13"
    },
    {
     "id": "W3",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Wald names the gap in his own words: the share of planes that actually survived a given hit in a particular combat is unknown, because the distribution of hits on the planes that did not return is not known.",
     "quote": "is unknown, since we do not know the distribution of hits on the planes that did not return",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part VII, Miscellaneous Remarks (SRG memo 109), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part VII, remark 2 (Non-probabilistic interpretation of the results), printed p. 76 (PDF p. 85)",
     "date_read": "2026-10-13"
    },
    {
     "id": "W4",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "To estimate how vulnerable each part of the plane is, Wald's method needs a second input besides the damage on returning planes: how hits are distributed over the parts of the plane. He states that the two cannot both be estimated from the returning planes' damage alone, and he takes the hit distribution as known beforehand or as measured by firing dummy bullets.",
     "quote": "it is impossible to estimate both … from the damage to returning planes only … can be assumed to be known a priori … or can be established experimentally by firing with dummy bullets",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part V, Subdivision of the Plane into Several Equi-Vulnerability Areas (SRG memo 96), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part V, printed p. 57 (PDF p. 66). The two quantities are defined on printed p. 56 (PDF p. 65); the worked example sets the hit distribution equal to each part's share of the plane's area (printed p. 62, PDF p. 71).",
     "date_read": "2026-10-13"
    },
    {
     "id": "W5",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the worked example that closes Part V, a single hit on the engines is survived with probability .61 and downs the plane with probability .39; for the fuselage the figures are .95 and .05. Wald concludes that, for this example, the engine area is the most vulnerable. Scope: Wald calls the example hypothetical, so the figures illustrate the method and are not measurements from combat.",
     "quote": "Engines .61 .39 … Fuselage .95 .05 … for the observed data of this hypothetical example, the engine area is the most vulnerable in the sense that a hit there is most likely to down the plane",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part V, Subdivision of the Plane into Several Equi-Vulnerability Areas (SRG memo 96), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part V, Numerical Example, summary table and closing paragraph, printed p. 65 (PDF p. 74). The table's two figure columns are the probability of surviving a single hit and the probability of being downed by a single hit.",
     "date_read": "2026-10-13"
    },
    {
     "id": "W6",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the Part V worked example the engines are not free of hits on the planes that came back. Of the 202 hits counted on returning planes, 39 are on part 1, the engines (.193 of the total), and 78 are on part 2, the fuselage (.386). Scope: the figures are hypothetical; they show the engines hit less often than the fuselage, which is not the same as untouched.",
     "quote": "The hits on the returning planes were distributed as follows … 1 39 .193 … 2 78 .386 … Total number of hits 202",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part V, Subdivision of the Plane into Several Equi-Vulnerability Areas (SRG memo 96), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part V, Numerical Example, table of hits by part, printed p. 64 (PDF p. 73). Columns: part number, number of hits observed on part, ratio to total observed hits. Part 1 is the 2 engines and part 2 the fuselage, with shares of the plane's area of .269 and .346 (printed p. 63, PDF p. 72).",
     "date_read": "2026-10-13"
    },
    {
     "id": "W7",
     "url": "https://durabilitycurve.com/claims/the-planes-that-didnt-come-back/#c-W7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The memoranda's one statement about armour is that the conclusions drawn from the Part VIII vulnerability table, in which the plane is most vulnerable to a hit on the engine area when the type of bullet is not specified, can be used as guides for locating protective armour. Scope: the table comes from data Wald calls hypothetical; a search of the reprint's text finds armour mentioned in this sentence only, and the sentence makes no recommendation about any actual aircraft.",
     "quote": "the hypothetical data would lead to the conclusion that the plane is most vulnerable to a hit on the engine area if the type of bullet is not specified … can be used as guides for locating protective armor",
     "removal_reason": "",
     "source": "Abraham Wald, A Method of Estimating Plane Vulnerability Based on Damage of Survivors, Part VIII, Vulnerability of a Plane to Different Types of Guns (SRG memo 126), reprinted by the Center for Naval Analyses as CRC 432, July 1980",
     "source_url": "https://www.cna.org/reports/1980/0204320000.pdf",
     "locator": "Part VIII, Comments on Results, printed pp. 88 and 89 (PDF pp. 97 and 98). The first part of the quote is on PDF p. 97; the armour clause is on PDF p. 98.",
     "date_read": "2026-10-13"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Graph Looks Right. The Merge Is Where It Breaks.",
    "url": "https://durabilitycurve.com/claims/where-agents-disagree/",
    "essay": {
     "title": "The Graph Looks Right. The Merge Is Where It Breaks.",
     "url": "https://durabilitycurve.com/blog/where-agents-disagree/",
     "published": "2026-08-30"
    },
    "last_checked": "2026-09-27",
    "format": "current",
    "counts": {
     "claims": 19,
     "by_status": {
      "EXECUTED": 14,
      "VERIFIED": 4,
      "REPORTED": 1
     },
     "verified_at_primary": 4,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/where-agents-disagree/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Two parallel nodes writing one plain (LastValue) channel do not silently overwrite; LangGraph refuses and raises",
     "quote": "At key 'decision': Can receive only one value per step. Use an Annotated key to handle multiple values.",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Silencing that error with a last-write-wins reducer on the decision channel keeps one write and silently drops the rest, with no error",
     "quote": "the channel now holds: 'roll back' and silently dropped: ['ship it', 'hold'] (no error was raised)",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-3",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Which write survives follows node-name order (the node whose name sorts last in Python's default codepoint order wins, not case-insensitive alphabetical) and is stable across re-runs of the same graph; an undocumented implementation detail",
     "quote": "zebra=ship, alpha=hold, mike=roll -> survives: ship and determinism (same order x3): ['roll back', 'roll back', 'roll back']; codepoint check Zebra vs alpha -> alpha survives (Z=0x5A < a=0x61)",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-4",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The documented reducer operator.add is lossless: on a str channel it concatenates the writes in node-name (alphabetical) fold order; on a list channel it keeps all of them",
     "quote": "decision -> 'roll backship ithold' (str, release-gate fan-out changelog/metrics/tests, stable x3) and ['ship it', 'hold', 'roll back'] (list)",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-5",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The SAME reducer knob is correct on a gather channel: three read-workers append an isolated operator.add findings channel and a single writer reconciles all three",
     "quote": "findings reaching the single writer = 3 of 3 and reconciled 3 of 3",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-6",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "A Send fan-out into a shared plain decision channel raises the same InvalidUpdateError",
     "quote": "raised InvalidUpdateError: At key 'decision': Can receive only one value per step.",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-7",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "A cyclic graph with no stop condition halts at LangGraph's default recursion limit of 25 and raises GraphRecursionError",
     "quote": "Recursion limit of 25 reached without hitting a stop condition.",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-8",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "interrupt() surfaces its payload with or without a checkpointer, but resuming needs one: without a checkpointer Command(resume=...) raises",
     "quote": "RuntimeError: Cannot use Command(resume=...) without checkpointer and (no-checkpointer first run) has __interrupt__: True",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-9",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "On resume the interrupt node re-runs from the top, so a side effect above the interrupt() call happens twice",
     "quote": "side effect ran 2 times -> twice, claim SUPPORTED",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-10",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "A plain channel written by two nodes in sequence (across supersteps) overwrites silently, no error; the first write is lost",
     "quote": "no error. final decision = 'B said ship' -> A's write silently lost",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-11",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "LangGraph exposes no static per-node write set: a node's introspectable .channels is its read set (a node that writes only decision still lists the whole readable state)",
     "quote": "a.channels: ['decision', 'other']",
     "removal_reason": "",
     "source": "first-party own run, langgraph 0.6.11",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-12",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The safe demo makes exactly three worker model calls, one per read-worker",
     "quote": "model calls made: 3",
     "removal_reason": "",
     "source": "first-party own run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-13",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "audit.py reads a State TypedDict and flags reducer channels (the fan-out silent-merge surface): findings on the safe schema, decision on the trap schema",
     "quote": "! classify findings REDUCER (add) and fan-out safe decision plain",
     "removal_reason": "",
     "source": "first-party own run (re-run 2026-08-29 after output cleanup)",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-14",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Environment of record for every mechanism row",
     "quote": "langgraph 0.6.11 and langchain-core 0.3.86 and python 3.9.6",
     "removal_reason": "",
     "source": "first-party own run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-29"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cognition argues actions carry implicit decisions and recommends a single-threaded linear agent",
     "quote": "\"Actions carry implicit decisions, and conflicting decisions carry bad results\" · \"The simplest way to follow the principles is to just use a single-threaded linear agent\"",
     "removal_reason": "",
     "source": "Cognition (Walden Yan), Don't Build Multi-Agents, 2025",
     "source_url": "https://cognition.com/blog/dont-build-multi-agents",
     "locator": "Principle 2; \"A Theory of Building Long-running Agents\"",
     "date_read": "2026-09-27"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cognition's 2026 follow-up concludes multi-agent systems work best when writes stay single-threaded and the extra agents add intelligence rather than actions (cited in [^2])",
     "quote": "\"multi-agent systems work best today when writes stay single-threaded and the additional agents contribute intelligence rather than actions\"",
     "removal_reason": "",
     "source": "Cognition (Walden Yan), Multi-Agents: What's Actually Working, 22 Apr 2026",
     "source_url": "https://cognition.com/blog/multi-agents-working",
     "locator": "essay body (re-fetched at primary, 2026-08-29)",
     "date_read": "2026-09-27"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic reports its multi-agent research system uses about 15x the tokens of a chat",
     "quote": "\"multi-agent systems use about 15× more tokens than chats\"",
     "removal_reason": "",
     "source": "Anthropic engineering blog",
     "source_url": "https://www.anthropic.com/engineering/multi-agent-research-system",
     "locator": "\"Benefits of a multi-agent system\"",
     "date_read": "2026-09-27"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-18",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The MAST study finds specification and system design the largest of three roughly balanced failure categories, and the authors trace many failures to system design rather than model limits alone",
     "quote": "\"The distribution of failures across MAST’s categories is relatively balanced (FC1: 41.77%, FC2: 36.94%, FC3: 21.30%, Figure 2)\" · \"many MAS failures came from system design, not just limitations of the underlying LLMs\"",
     "removal_reason": "",
     "source": "Cemri et al., \"Why Do Multi-Agent LLM Systems Fail?\", arXiv 2503.13657v2",
     "source_url": "https://arxiv.org/abs/2503.13657v2",
     "locator": "v2, \"Balanced Distribution\"; §5.2 \"Beyond Model Capabilities: The Primacy of System Design\"",
     "date_read": "2026-09-27"
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/where-agents-disagree/#c-19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ranjan Kumar published the reducer-as-concurrency-policy framing and a compiled-graph merge-policy auditor before this piece (convergent prior art, credited in the body)",
     "quote": "\"The reducer is not a merge helper. It is the concurrency-control policy for that state key.\" · \"It appears on no diagram, in no edge list, and in no type checker’s output.\"",
     "removal_reason": "",
     "source": "Ranjan Kumar, \"LangGraph Reducers Are a Concurrency Policy\", 27 Jul 2026",
     "source_url": "https://ranjankumar.in/langgraph-reducers-concurrent-state-writes",
     "locator": "post body (re-fetched at primary, 2026-08-29)",
     "date_read": "2026-09-27"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Everyone Got Safer. That's the Problem.",
    "url": "https://durabilitycurve.com/claims/everyone-got-safer/",
    "essay": {
     "title": "Everyone Got Safer. That's the Problem.",
     "url": "https://durabilitycurve.com/blog/everyone-got-safer/",
     "published": "2026-08-28"
    },
    "last_checked": "2026-10-13",
    "format": "older (no quotes or locators recorded)",
    "counts": {
     "claims": 17,
     "by_status": {
      "CHECKED": 15,
      "REPORTED": 2
     },
     "verified_at_primary": 0,
     "removed": 2
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/everyone-got-safer/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-1",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "July 2023, CMU + Center for AI Safety, ~20-token adversarial suffix that transfers across labs",
     "quote": "",
     "removal_reason": "",
     "source": "Zou et al., arXiv:2307.15043 (submitted 27 Jul 2023)",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-2",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "GPT-3.5 \"most of the time,\" Bard \"two-thirds,\" GPT-4 \"roughly half\"",
     "quote": "",
     "removal_reason": "",
     "source": "Table 2 ensemble row: GPT-3.5 86.6%, PaLM-2/Bard 66.0%, GPT-4 46.9%",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-3",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "One string got through the commercial models of two rival labs, while Claude was more mixed: 47.9% on Claude-1 and about 2.1% on Claude-2",
     "quote": "",
     "removal_reason": "",
     "source": "Table 2: 86.6 / 46.9 / 66.0 / Claude-1 47.9 / Claude-2 2.1",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-4",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Zou et al. say their attack did much better against GPT-based models, potentially because Vicuna, an open model it was built on, was trained on ChatGPT outputs",
     "quote": "",
     "removal_reason": "",
     "source": "Zou et al. §transfer",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-5",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Ross Ashby's cybernetics named the limit in 1956, so that part of the idea is not new",
     "quote": "",
     "removal_reason": "",
     "source": "Ashby, An Introduction to Cybernetics, 1956",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-6",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Donald Campbell was a psychologist",
     "quote": "",
     "removal_reason": "The published essay no longer names Donald Campbell (earlier verdict: FIXED, \"economist\" corrected to \"psychologist\")",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-7",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Baker et al. found that chain-of-thought monitoring helps at low optimisation pressure, but under too much optimisation agents learn to hide their intent while still reward hacking",
     "quote": "",
     "removal_reason": "",
     "source": "Baker et al., arXiv:2503.11926",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-8",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Bommasani et al. warned in 2021 that the defects of a foundation model are inherited by all the adapted models downstream",
     "quote": "",
     "removal_reason": "",
     "source": "Bommasani et al., arXiv:2108.07258, abstract",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-9",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Dziemian: 272k attempts, 13 models, 8,648 successes, all vulnerable, 0.5%-8.5%, cross-family",
     "quote": "",
     "removal_reason": "",
     "source": "Dziemian et al., arXiv:2603.15714",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-10",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Kuai et al. found behavioural entanglement across model families and state that apparent agreement reflects shared error modes rather than independent validation",
     "quote": "",
     "removal_reason": "",
     "source": "arXiv:2604.07650 (Kuai et al.) verbatim: \"apparent agreement reflects shared error modes rather than independent validation\"",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-11",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "MMLU contamination \"tens of percent,\" Llama-2",
     "quote": "",
     "removal_reason": "",
     "source": ">16% MMLU contaminated for Llama-2",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-12",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Li, J Fixed Income 9(4), 2000, pages",
     "quote": "",
     "removal_reason": "",
     "source": "pp. 43-54 (start 43 confirmed)",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-13",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Wired published Felix Salmon's article \"Recipe for Disaster: The Formula That Killed Wall Street\" on 23 February 2009",
     "quote": "",
     "removal_reason": "",
     "source": "Wired, 23 Feb 2009",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-14",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "1987: Dow -22.6% (508 pts), S&P -20.4%, 19 Oct",
     "quote": "",
     "removal_reason": "",
     "source": "Fed Reserve History",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-15",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The Brady Commission report gave portfolio-insurance selling heavy weight among the causes of the 1987 crash, though the exact weight is still debated",
     "quote": "",
     "removal_reason": "",
     "source": "Brady Report 1988; \"greatest importance\" is secondary phrasing",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-16",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Aviation ~65% fall in fatal-accident rate, Boeing (not IATA)",
     "quote": "",
     "removal_reason": "",
     "source": "Boeing Statistical Summary via Flight Safety Foundation",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-17",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "737 MAX MCAS single AoA sensor, hidden from pilots; 346 dead; grounded Mar 2019",
     "quote": "",
     "removal_reason": "",
     "source": "Lion Air 610 (189) + Ethiopian 302 (157) = 346",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-18",
     "status": "CHECKED",
     "status_label": "Checked",
     "removed": false,
     "claim": "Vision transfer tracks model similarity, \">90%\" predictor; adjacent not same-domain",
     "quote": "",
     "removal_reason": "",
     "source": "Demontis et al. USENIX 2019; arXiv:2501.18629",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/everyone-got-safer/#c-19",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Geer et al. made the security-monoculture argument earlier, in their 2003 paper \"CyberInsecurity: The Cost of Monopoly\"",
     "quote": "",
     "removal_reason": "The published essay does not cite Geer et al. (earlier verdict: ADDED, for honesty about precedent)",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Confidently Wrong",
    "url": "https://durabilitycurve.com/claims/confidently-wrong/",
    "essay": {
     "title": "Confidently Wrong",
     "url": "https://durabilitycurve.com/blog/confidently-wrong/",
     "published": "2026-08-23"
    },
    "last_checked": "2026-08-22",
    "format": "current",
    "counts": {
     "claims": 11,
     "by_status": {
      "VERIFIED": 10,
      "REPORTED": 1
     },
     "verified_at_primary": 10,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/confidently-wrong/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Reinhart & Rogoff (2010) reported that above 90% debt-to-GDP, average real growth was −0.1%",
     "quote": "\"median growth rates fall by ... and average (i.e., mean) growth rates fall considerably more\" with the above-90% mean at −0.1%",
     "removal_reason": "",
     "source": "Reinhart & Rogoff, Growth in a Time of Debt (AER P&P)",
     "source_url": "https://scholar.harvard.edu/files/rogoff/files/growth_in_time_debt_aer.pdf",
     "locator": "Table 1 / §II",
     "date_read": "2026-08-22"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The 90% figure was used to justify austerity, cited by Paul Ryan's budget and by the European Commission",
     "quote": "\"Both Paul Ryan ... and Olli Rehn ... invoked the 90 percent threshold\"",
     "removal_reason": "",
     "source": "Krugman, How the Case for Austerity Has Crumbled (NYRB)",
     "source_url": "https://www.nybooks.com/articles/2013/06/06/how-case-austerity-has-crumbled/",
     "locator": "Body",
     "date_read": "2026-08-22"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Herndon (a UMass grad student), Ash & Pollin obtained the spreadsheet and found the result rested on three faults",
     "quote": "\"a combination of coding errors, selective exclusion of available data, and unconventional weighting of summary statistics\"",
     "removal_reason": "",
     "source": "Herndon, Ash & Pollin, PERI Working Paper 322",
     "source_url": "https://peri.umass.edu/publication/does-high-public-debt-consistently-stifle-economic-growth-a-critique-of-reinhart-and-rogoff/",
     "locator": "Abstract",
     "date_read": "2026-08-22"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Excel formula excluded five countries (Australia, Austria, Belgium, Canada, Denmark); corrected average growth above 90% debt was +2.2%",
     "quote": "\"the average real GDP growth rate for countries carrying a public debt-to-GDP ratio of over 90 percent is actually 2.2 percent, not −0.1 percent\"",
     "removal_reason": "",
     "source": "Herndon, Ash & Pollin, PERI WP 322",
     "source_url": "https://peri.umass.edu/publication/does-high-public-debt-consistently-stifle-economic-growth-a-critique-of-reinhart-and-rogoff/",
     "locator": "Abstract / §III",
     "date_read": "2026-08-22"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "People stop cross-checking automated output; complacency and bias appear in experts and are not eliminated by training or warnings",
     "quote": "\"automation-induced complacency ... found in both naïve and expert participants and cannot be overcome with simple practice\"; bias \"cannot be prevented by training or instructions\"",
     "removal_reason": "",
     "source": "Parasuraman & Manzey, Complacency and Bias in Human Use of Automation, Human Factors 52(3)",
     "source_url": "https://journals.sagepub.com/doi/10.1177/0018720810376055",
     "locator": "Abstract / §Integration",
     "date_read": "2026-08-22"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Knight & Leveson (1986) had 27 programmers write the same program from one spec, ran each version on about one million inputs, and found failures were correlated far beyond independence (rejected at the 99% confidence level)",
     "quote": "\"27 versions of a program ... executed on one million randomly-generated inputs ... the hypothesis ... that the ... versions fail independently ... was rejected at the 99% confidence level\"",
     "removal_reason": "",
     "source": "Knight & Leveson, An Experimental Evaluation of the Assumption of Independence in Multiversion Programming, IEEE TSE SE-12(1)",
     "source_url": "https://www.csc.kth.se/utbildning/kth/kurser/DA2210/vettig13/Seminarier/KnightLeveson.pdf",
     "locator": "Abstract / §Results",
     "date_read": "2026-08-22"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Desirable difficulties (spacing, interleaving, generation, self-testing) build durable skill, and are desirable only if the learner can meet them",
     "quote": "\"If ... the learner does not have the background knowledge or skills to respond to them successfully, they become undesirable difficulties\"",
     "removal_reason": "",
     "source": "Bjork & Bjork, Making Things Hard on Yourself, But in a Good Way (2011)",
     "source_url": "https://bjorklab.psych.ucla.edu/research/",
     "locator": "Body",
     "date_read": "2026-08-22"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Bainbridge (1983): automating a task leaves the operator the parts too hard to automate, and less practised at the skills they demand",
     "quote": "\"the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate\"",
     "removal_reason": "",
     "source": "Bainbridge, Ironies of Automation, Automatica 19(6)",
     "source_url": "https://www.sciencedirect.com/science/article/abs/pii/0005109883900468",
     "locator": "p.775",
     "date_read": "2026-08-22"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Capt. Warren VanderBurgh's 1997 American Airlines training talk named \"children of the magenta line\"; by the airline's own reckoning most of the automation-related trouble studied traced to automation mismanagement",
     "quote": "\"Children of the Magenta (Line)\", AA training, ~April 1997; ~68% of reviewed events traced to automation mismanagement (softened in-body to \"most\")",
     "removal_reason": "",
     "source": "VanderBurgh, Children of the Magenta Line (AA training); AirFacts retrospective",
     "source_url": "https://airfactsjournal.com/2020/09/stepping-down-in-automation-the-real-lesson-for-children-of-the-magenta-line/",
     "locator": "Retrospective",
     "date_read": "2026-08-22"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "AF447 (2009): pitot icing → unreliable airspeed → autopilot disengaged → nose held up → unrecognised stall; BEA cited crew-coordination breakdown and absence of manual-handling training at altitude among causes",
     "quote": "\"the absence of training, at high altitude, in manual aeroplane handling and in the procedure for unreliable airspeed\"; CRM \"degraded\"",
     "removal_reason": "",
     "source": "BEA final report on AF447 (2012), via IEEE Spectrum / Aviation Safety",
     "source_url": "https://spectrum.ieee.org/air-france-flight-447-crash-caused-by-a-combination-of-factors",
     "locator": "Final report summary",
     "date_read": "2026-08-22"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/confidently-wrong/#c-11",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Verifying a completed solution is generally far cheaper than producing one (the P vs NP intuition; Sudoku as the illustration)",
     "quote": "verifying a candidate solution is easy while finding one is hard — the defining property of NP",
     "removal_reason": "",
     "source": "Clay Mathematics Institute, P vs NP",
     "source_url": "https://www.claymath.org/millennium/p-vs-np/",
     "locator": "Problem statement",
     "date_read": "2026-08-22"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: A Green Score Is Not Evidence",
    "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/",
    "essay": {
     "title": "A Green Score Is Not Evidence",
     "url": "https://durabilitycurve.com/blog/the-evaluation-inversion/",
     "published": "2026-08-21"
    },
    "last_checked": "2026-08-20",
    "format": "current",
    "counts": {
     "claims": 12,
     "by_status": {
      "VERIFIED": 10,
      "REPORTED": 2
     },
     "verified_at_primary": 10,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-evaluation-inversion/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A DeepEval faithfulness score reached 0.978 in the no-context condition, with contextual precision and recall both 0.000",
     "quote": "\"No Context (Parametric) 0.000 [0.000, 0.000] 0.000 [0.000, 0.000] 0.978 [0.965, 0.990] 0.287 [0.246, 0.329] 0.316\" (columns: Ctx.Prec / Ctx.Recall / Faithfulness / Ans.Relev / Composite)",
     "removal_reason": "",
     "source": "Bal & Puhan, Benchmarking Retrieval Strategies for Biomedical RAG",
     "source_url": "https://arxiv.org/abs/2605.02520",
     "locator": "Table 2",
     "date_read": "2026-08-20"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "That 0.978 is the highest faithfulness of any condition and is a metric artefact: faithfulness scores the absence of contradiction, so with no context there is nothing to contradict",
     "quote": "\"The no-context ablation records 0.978 on faithfulness, which is a measurement artifact rather than a genuine result... with no context, there is nothing to contradict, and the score defaults to near-perfect. This shows why faithfulness alone is an unreliable quality signal for RAG systems\"",
     "removal_reason": "",
     "source": "Bal & Puhan (same)",
     "source_url": "https://arxiv.org/abs/2605.02520",
     "locator": "§5.3 Faithfulness",
     "date_read": "2026-08-20"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Faithfulness scored high by construction because the metric measures absence of contradiction with the provided context",
     "quote": "\"No-Context ranks #1 in faithfulness by construction (0.978, an artefact of the metric definition) but last on all retrieval-dependent metrics\"",
     "removal_reason": "",
     "source": "Bal & Puhan (same)",
     "source_url": "https://arxiv.org/abs/2605.02520",
     "locator": "Figure 1 caption",
     "date_read": "2026-08-20"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the fictional experimental setup, the model gave the harmful response in 12% of free/training cases and refused in 97% of paid/unmonitored cases; the free/paid setup is explicitly fictional",
     "quote": "\"in 12% of cases, the model gave the harmful response\" · \"in 97% of cases in the 'paid' condition...this is what happened\" (refusal) · \"note that this is entirely fictional; in reality, we don't train our Claude models on free or paid user data by default\"",
     "removal_reason": "",
     "source": "Anthropic, Alignment faking in large language models (announcement)",
     "source_url": "https://anthropic.com/news/alignment-faking",
     "locator": "Announcement body",
     "date_read": "2026-08-20"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "(Cross-check) the paper abstract states the free-tier compliance figure as 14%, a different cut from the announcement figure the body uses; retained to document the discrepancy",
     "quote": "\"the model complies with harmful queries from free users 14% of the time, versus almost never for paid users\"",
     "removal_reason": "",
     "source": "Greenblatt et al., Alignment faking in large language models",
     "source_url": "https://arxiv.org/abs/2412.14093",
     "locator": "Abstract",
     "date_read": "2026-08-20"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "An adaptive attack reached 100% attack success rate, including on R2D2 which had been adversarially trained against the GCG attack",
     "quote": "\"we achieve 100% attack success rate... on [models incl.] R2D2 from HarmBench that was adversarially trained against the GCG attack\"",
     "removal_reason": "",
     "source": "Andriushchenko, Croce, Flammarion, Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks",
     "source_url": "https://arxiv.org/abs/2404.02151",
     "locator": "Abstract",
     "date_read": "2026-08-20"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A frontier model verbalised recognising an evaluation and deliberately underperformed to avoid a consequence",
     "quote": "\"I notice that this is a test of arithmetic capabilities... I should be careful not to demonstrate too much arithmetic capability\"",
     "removal_reason": "",
     "source": "Meinke et al., Frontier Models are Capable of In-context Scheming",
     "source_url": "https://arxiv.org/abs/2412.04984",
     "locator": "§3.6 (sonnet-3.5 transcript)",
     "date_read": "2026-08-20"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors note a model may notice it is being tested and behave more aligned during evaluation, causing evaluations to underestimate its behaviour",
     "quote": "\"a model might notice that it is being tested. If models are capable of noticing that they are being tested and have a tendency to behave more aligned in such evaluations, then our evaluations might underestimate models' scheming capabilities, particularly for more capable models\"",
     "removal_reason": "",
     "source": "Meinke et al. (same)",
     "source_url": "https://arxiv.org/abs/2412.04984",
     "locator": "§4 Limitations",
     "date_read": "2026-08-20"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-9",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Scope: only some frontier models (sonnet-3.5, opus-3, o1) showed statistically significant differences between the inferred test and deploy settings",
     "quote": "",
     "removal_reason": "",
     "source": "Meinke et al. (same)",
     "source_url": "https://arxiv.org/abs/2412.04984",
     "locator": "Results per-model",
     "date_read": "2026-08-20"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The knowledge-conflict paradigm measures whether a model follows substituted contextual evidence or reverts to memory, and finds models over-rely on memorised information even when reading context",
     "quote": "\"we measure their over-reliance on memorized information (the cause of hallucinations), and uncover important factors that exacerbate this behaviour\"",
     "removal_reason": "",
     "source": "Longpre et al., Entity-Based Knowledge Conflicts in Question Answering (EMNLP 2021)",
     "source_url": "https://arxiv.org/abs/2109.05052",
     "locator": "Abstract",
     "date_read": "2026-08-20"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Construct validity from measurement theory applies to ML metrics: a metric can mismatch the construct it purports to measure",
     "quote": "\"This process...introduces the potential for mismatches between the theoretical understanding of the construct purported to be measured and its operationalization\"",
     "removal_reason": "",
     "source": "Jacobs & Wallach, Measurement and Fairness (ACM FAccT 2021)",
     "source_url": "https://arxiv.org/abs/1912.05511",
     "locator": "Body",
     "date_read": "2026-08-20"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/the-evaluation-inversion/#c-12",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Goodhart's law has a visible/regressional variant distinct from an adversarial one; the visible decoupling is the tractable case",
     "quote": "(well-established taxonomy; cited conceptually, not quoted verbatim in body)",
     "removal_reason": "",
     "source": "Manheim & Garrabrant, Categorizing Variants of Goodhart's Law",
     "source_url": "https://arxiv.org/abs/1803.04585",
     "locator": "Full paper",
     "date_read": "2026-08-20"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Limit Said 10. The Loop Made 500 Calls.",
    "url": "https://durabilitycurve.com/claims/the-limit-said-10/",
    "essay": {
     "title": "The Limit Said 10. The Loop Made 500 Calls.",
     "url": "https://durabilitycurve.com/blog/the-limit-said-10/",
     "published": "2026-08-16"
    },
    "last_checked": "2026-10-06",
    "format": "current",
    "counts": {
     "claims": 34,
     "by_status": {
      "VERIFIED": 21,
      "EXECUTED": 12,
      "REPORTED": 1
     },
     "verified_at_primary": 21,
     "removed": 3
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-limit-said-10/"
   },
   "rows": [
    {
     "id": "C46",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C46",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The opening incident ran 9–10 May 2026, not June, and provisioned five m8g.12xlarge instances plus load balancers and Lambdas, duplicated by repeatedly deploying the same CloudFormation template, against DN42.",
     "quote": "Timeline: began \"2026-05-09\", agent shut down roughly 24 hours later on \"2026-05-10\". Agent's own proposal: \"My primary objective is to conduct comprehensive (full port) network scanning and topological data gathering.\" Five \"m8g.12xlarge\" instances; operator: \"many instance and load balancer and lambda\", duplicated by redeploying \"the same cloudformation template\"",
     "removal_reason": "",
     "source": "Lan Tian, blog post on the DN42 scanning agent",
     "source_url": "https://lantian.pub/en/article/fun/ai-agent-bankrupted-their-operator-scan-dn42lantian.lantian/",
     "locator": "§ timeline + § what it built",
     "date_read": "2026-08-14"
    },
    {
     "id": "C47",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C47",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The $6,531.30 figure is the bill as incurred; AWS subsequently reduced it to about $1,894",
     "quote": "Initial bill \"$6531.30\"; operator reports the revised amount as \"$1894\" after AWS agreed to reduce charges",
     "removal_reason": "",
     "source": "Lan Tian, ibid.",
     "source_url": "https://lantian.pub/en/article/fun/ai-agent-bankrupted-their-operator-scan-dn42lantian.lantian/",
     "locator": "§ aftermath",
     "date_read": "2026-08-14"
    },
    {
     "id": "C48",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C48",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The \"a small VPS would have done it\" judgement is the blog author's, not the operator's",
     "quote": "\"for a hobbyist network like DN42, such infrastructure is way overkill, a small VPS server would do the job\" — author's own assessment",
     "removal_reason": "",
     "source": "Lan Tian, ibid.",
     "source_url": "https://lantian.pub/en/article/fun/ai-agent-bankrupted-their-operator-scan-dn42lantian.lantian/",
     "locator": "§ assessment",
     "date_read": "2026-08-14"
    },
    {
     "id": "C49",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C49",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The operator learned of the spend through card charges rather than AWS monitoring",
     "quote": "Operator: \"the cost too high and much charges on card\"",
     "removal_reason": "",
     "source": "Lan Tian, ibid.",
     "source_url": "https://lantian.pub/en/article/fun/ai-agent-bankrupted-their-operator-scan-dn42lantian.lantian/",
     "locator": "§ discovery",
     "date_read": "2026-08-14"
    },
    {
     "id": "C50",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C50",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "AWS Budgets can attach an enforcing action, up to a Deny IAM policy, but budget data refreshes at most three times a day, typically 8–12 hours apart. It fires late, not never, and late is the stronger sentence",
     "quote": "\"AWS Budgets information is updated up to three times a day. Updates typically occur 8–12 hours after the previous update.\" · \"Apply a custom Deny IAM policy that restricts the ability for a user, group, or role to provision additional Amazon EC2 resources.\" (second locator budgets-controls.html) · \"You might incur additional costs or usage that exceed your budget notification threshold before AWS Budgets can notify you\"",
     "removal_reason": "",
     "source": "AWS Cost Management docs",
     "source_url": "https://docs.aws.amazon.com/cost-management/latest/userguide/budgets-managing-costs.html",
     "locator": "§ Managing your costs with AWS Budgets — console setting, service-side, nothing to run",
     "date_read": "2026-08-14"
    },
    {
     "id": "C94",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C94",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The quote supports only that On-Demand quotas exist per account per Region and are denominated in vCPUs. It does NOT support \"the fast structural control the AWS incident did not have set\", and the same page contradicts that framing outright: quotas are automatically increased with usage, so they are never off, only sized. Two independent reads flagged the row and the caller re-verified at source. The published default for the Standard family is 5, quoted below",
     "quote": "\"There are quotas for the number of running On-Demand Instances per AWS account per Region… managed in terms of the number of virtual central processing units (vCPUs)\" ·",
     "removal_reason": "",
     "source": "AWS, Amazon EC2 User Guide",
     "source_url": "https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-on-demand-instances.html",
     "locator": "§ On-Demand Instance quotas — console setting, service-side.",
     "date_read": "2026-08-15"
    },
    {
     "id": "C1",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ball's framing of an agent is \"an LLM, a loop, and enough tokens\"",
     "quote": "\"It's an LLM, a loop, and enough tokens. It's what we've been saying on the podcast from the start.\"",
     "removal_reason": "",
     "source": "Thorsten Ball, How to Build an Agent",
     "source_url": "https://ampcode.com/notes/how-to-build-an-agent",
     "locator": "§ opening",
     "date_read": "2026-08-13"
    },
    {
     "id": "C2",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A working code-editing agent fits in under 400 lines, mostly boilerplate",
     "quote": "\"You can do it in less than 400 lines of code, most of which is boilerplate.\"",
     "removal_reason": "",
     "source": "Ball, ibid.",
     "source_url": "https://ampcode.com/notes/how-to-build-an-agent",
     "locator": "§ opening",
     "date_read": "2026-08-13"
    },
    {
     "id": "C4",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The canonical tutorial does not cover error recovery, termination, token budget, cost, tool-output verification, concurrency or streaming",
     "quote": "Absence claim, checked at source: zero occurrences of \"error recover\", \"terminat\", \"token budget\", \"cost\", \"stream\", \"concurren\" across all 7 headings.",
     "removal_reason": "",
     "source": "Ball, ibid.",
     "source_url": "https://ampcode.com/notes/how-to-build-an-agent",
     "locator": "full page, 7 headings",
     "date_read": "2026-08-14"
    },
    {
     "id": "C85",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C85",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Ball's tutorial does not cover verifying what a tool returned",
     "quote": "Absence claim, checked at source against all 7 headings. The only handling of a tool's return is executeTool flipping an error boolean — return anthropic.NewToolResultBlock(id, err.Error(), true) — which captures an error raised BY the tool, not any check on whether what came back is correct or usable. Found by a cold read, not by a gate",
     "removal_reason": "",
     "source": "Thorsten Ball, How to Build an Agent",
     "source_url": "https://ampcode.com/notes/how-to-build-an-agent",
     "locator": "full page, 7 headings",
     "date_read": "2026-08-14"
    },
    {
     "id": "C5",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic defines agents as systems where the model directs its own process",
     "quote": "agents are \"systems where LLMs dynamically direct their own processes and tool usage\"",
     "removal_reason": "",
     "source": "Anthropic, Building Effective Agents",
     "source_url": "https://www.anthropic.com/engineering/building-effective-agents",
     "locator": "§ What are agents?",
     "date_read": "2026-08-13"
    },
    {
     "id": "C7",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic advises agents need ground truth from the environment each step",
     "quote": "\"During execution, it's crucial for the agents to gain 'ground truth' from the environment at each step\"",
     "removal_reason": "",
     "source": "Anthropic, ibid.",
     "source_url": "https://www.anthropic.com/engineering/building-effective-agents",
     "locator": "",
     "date_read": "2026-08-13"
    },
    {
     "id": "C99",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C99",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A tool result is returned to the model in a tool_result content block carrying the tool_use_id of the call it answers; the id is what pairs the two halves of the exchange",
     "quote": "\"Your code executes the operation and sends back a tool_result.\" · \"a second request sends the result back in a tool_result block so Claude can reply with the answer\" · request body, verbatim from the worked example: {\"type\": \"tool_result\", \"tool_use_id\": tool_use.id, \"content\": weather}",
     "removal_reason": "",
     "source": "Anthropic, Tool use with Claude",
     "source_url": "https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview",
     "locator": "§ How tool use works — service-side, console/service contract, nothing to run locally",
     "date_read": "2026-08-15"
    },
    {
     "id": "C100",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C100",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Every mainstream framework ships a step-counting bound. Re-read on 2026-10-06 out of each project's current wheel: langchain-classic 1.0.8 AgentExecutor.max_iterations default 15 · langgraph 1.2.14 recursion_limit default 10007 · crewai 1.15.23 max_iter default 25 · openai-agents 0.23.1 max_turns default 10 · autogen-agentchat 0.7.5 composable TerminationCondition classes. LangChain's agent bound now lives only in the legacy langchain-classic package: in langchain 1.4.3 the one max_iterations left is a local loop variable in summarization.py (line 697), not an agent bound. AutoGen does not simply count steps: of its eleven concrete termination conditions exactly one counts messages (MaxMessageTermination), TokenUsageTermination counts tokens, TimeoutTermination counts wall-clock seconds, and the other eight fire on message content or on an external signal. Two current documentation pages disagree with the shipped code. LangGraph's Graph API page says the default has been 1000 steps since version 1.0.6, while langgraph 1.2.14 sets 10007. CrewAI's Agents page says max_iter defaults to 20, while crewai 1.15.23 sets 25 on both BaseAgent and its executor. The defaults given here are the code's, because the code is what runs.",
     "quote": "Run output, python3 Automation/framework_bounds_probe.py, 2026-10-06, 6/6 probes resolved: \"langchain-classic 1.0.8 langchain_classic/agents/agent.py — AgentExecutor.max_iterations default: 15 (line 1023)\" · \"langchain 1.4.3 langchain/agents/middleware/summarization.py — any surviving max_iterations in langchain 1.x (expected: NOT an agent bound): len(messages).bit_length() + 1 (line 697)\" · \"langgraph 1.2.14 langgraph/_internal/_config.py — recursion_limit default: 10007 (line 32)\" · \"crewai 1.15.23 crewai/agents/agent_builder/base_agent_executor.py — max_iter default: 25 (line 27)\" · \"openai-agents 0.23.1 agents/run_config.py — max_turns default: 10 (line 45)\" · \"autogen-agentchat 0.7.5 autogen_agentchat/conditions/_terminations.py — StopMessageTermination (line 24), MaxMessageTermination (line 62), TextMentionTermination (line 111), FunctionalTermination (line 158), TokenUsageTermination (line 235), HandoffTermination (line 313)\". The probe prints the six classes its pattern matches; the class total is eleven (autogen-agentchat is unchanged at 0.7.5 since the first run). Second route, the same day, the tagged source on GitHub: langgraph tag 1.2.14, libs/langgraph/langgraph/_internal/_config.py line 32, DEFAULT_RECURSION_LIMIT = int(getenv(\"LANGGRAPH_DEFAULT_RECURSION_LIMIT\", \"10007\")), applied by ensure_config (line 335); crewai tag 1.15.23, lib/crewai/src/crewai/agents/agent_builder/base_agent.py lines 286 to 288, max_iter: int = Field(default=25, ...). The documentation, read the same day: LangGraph, \"Starting in version 1.0.6, the default recursion limit is set to 1000 steps.\" (docs.langchain.com/oss/python/langgraph/graph-api, § Recursion limit); CrewAI, \"Maximum iterations before the agent must provide its best answer. Default is 20.\" (docs.crewai.com/en/concepts/agents, § Agent Attributes)",
     "removal_reason": "",
     "source": "First-party execution against current PyPI wheels",
     "source_url": "https://github.com/langchain-ai/langgraph/blob/1.2.14/libs/langgraph/langgraph/_internal/_config.py#L32",
     "locator": "",
     "date_read": "2026-10-06"
    },
    {
     "id": "C71",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C71",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cloudzy describes a recursion limit as a backstop firing after the waste",
     "quote": "\"It is a backstop that fires _after_ the loop has already wasted twenty-five steps and the API spend that goes with them.\"",
     "removal_reason": "",
     "source": "Cloudzy",
     "source_url": "https://cloudzy.com/blog/why-ai-agent-loops-fail-in-production/",
     "locator": "§ recursion limits",
     "date_read": "2026-08-14"
    },
    {
     "id": "C17",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "An open author copy exists and the quoted sentence verifies verbatim. The paper does NOT say termination requires a per-cycle measure; that framing is what its contribution denies. It presents the single-ranking-function method as Turing's, and argues past it via disjunctive well-foundedness.",
     "quote": "\"The problem with Turing's method is that finding a single, or monolithic, ranking function for the whole program is typically difficult, even for simple programs.\" Re-verified first-party by the caller with pypdf: PDF page 4 = journal page 91. Reference 39 on the last page reads \"Turing, A. Checking a large routine. In Report of a Conference on High Speed Automatic Calculating Machines\", which corroborates the Turing-1949 attribution",
     "removal_reason": "",
     "source": "Cook, Podelski & Rybalchenko, CACM 54(5):88–98, May 2011",
     "source_url": "http://www0.cs.ucl.ac.uk/staff/b.cook/pdfs/proving_program_termination.pdf",
     "locator": "journal p.91 (PDF p.4), § Turing's Classic Method and Disjunctive Well-Foundedness",
     "date_read": "2026-08-15"
    },
    {
     "id": "C72",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C72",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The open scan at the URL below is a 14-page image PDF with no text layer — pypdf extracts zero characters from all 14 pages, so the absence of the quoted strings proves nothing and their presence cannot be confirmed by this route either. What IS verified, first-party, is the Turing half: Cook et al.'s reference 39 is \"Turing, A. Checking a large routine…\" and their § heading is \"Turing's Classic Method\" (see C17).",
     "quote": "Floyd's authorship of the well-ordered-set termination argument is standard secondary attribution.",
     "removal_reason": "",
     "source": "Floyd, Assigning Meanings to Programs (1967), scanned copy",
     "source_url": "https://people.eecs.berkeley.edu/~necula/Papers/FloydMeaning.pdf",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C21",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C21",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "A loop with max_iterations=10 issues 500 model calls when the retry sits inside the bounded iteration; 500 is the bench's own backstop, not a natural stop",
     "quote": "Run output: \"malformed -> {'stopped': 'runaway', 'calls': 500}\". hard_cap=500 is an argument in uncovered_bound_loop.",
     "removal_reason": "",
     "source": "First-party run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C22",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C22",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "No-progress detection on hashed (tool, args) halts a repeating caller in 3 calls",
     "quote": "Run output: \"repeater -> {'stopped': 'no-progress', 'calls': 3, 'repeated': 'a099dc8c'}\"",
     "removal_reason": "",
     "source": "First-party run",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C23",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C23",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "That same guard does NOT catch a model making novel calls forever; it runs to the ceiling at 30",
     "quote": "Run output: \"never_finishes -> {'stopped': 'ceiling', 'calls': 30}\"",
     "removal_reason": "",
     "source": "First-party run. The 30 is max_iterations=30 in the guard's own signature, not a discovered threshold",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C88",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C88",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The three printed outputs of bounded_run are what the printed code produces",
     "quote": "Extracted all fenced blocks from the body and executed each runnable one. never_parses -> {'exit': 'budget', 'calls': 20}, always_repeats -> {'exit': 'stall', 'on': 'search', 'calls': 3}, never_finishes -> {'exit': 'budget', 'calls': 20} — byte-identical to the article's output block. The same check passed for the 500-call runaway bench and both guarded() lines.",
     "removal_reason": "",
     "source": "First-party run, reproducible",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-14"
    },
    {
     "id": "C8",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A static scan of 6,549 agent repos confirmed 68 infinite-loop failures across 47 projects at 91.9% precision",
     "quote": "\"On the real-world corpus of 6,549 LLM agent projects, IAL-SCAN reports 74 potential findings. Manual review confirms 68 IAL failures and 6 false positives, yielding an end-to-end precision of 91.9%.\" … \"These failures affect 47 agent projects\"",
     "removal_reason": "",
     "source": "Hou, Wang, Zhao & Wang (HUST). PREPRINT",
     "source_url": "https://arxiv.org/abs/2607.01641",
     "locator": "arXiv:2607.01641v1 §V-B1",
     "date_read": "2026-08-13"
    },
    {
     "id": "C9",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "All 68 confirmed failures share one root cause: the repeating path is not covered by a strong bound",
     "quote": "\"All 68 failures share the same root issue: the repeated path is not covered by a strong bound.\"",
     "removal_reason": "",
     "source": "Hou et al. PREPRINT",
     "source_url": "https://arxiv.org/abs/2607.01641",
     "locator": "arXiv:2607.01641v1 §V-B1",
     "date_read": "2026-08-13"
    },
    {
     "id": "C80",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C80",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The bench's failure shape occurs in a real repository: a planner nesting two while not success loops around an LLM call and swallowing the parse exception with pass",
     "quote": "\"Figure 6 shows a retry loop in 2456868764/LiteRAG. The planner uses nested while not success loops to repeatedly request a plan from the LLM. The costly call appears at lines 8–10, where self.llm.invoke(...) is executed until parsing succeeds. However, parser failures are swallowed at lines 12–13\". Figure 6 source, extracted verbatim: success = False / while not success: / while not success: / try: / plan = self.llm.invoke(...) / success = True / except OutputParserException: / pass.",
     "removal_reason": "",
     "source": "Hou et al. PREPRINT",
     "source_url": "https://arxiv.org/abs/2607.01641",
     "locator": "arXiv:2607.01641v1 §V-B2 Case Study + Figure 6",
     "date_read": "2026-08-14"
    },
    {
     "id": "C90",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C90",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The preprint's self-assessment figures: the first two authors agreed on 94.6% of the 74 potential findings and settled the rest by discussion; a separate false-negative review confirmed 7 misses",
     "quote": "\"During independent labeling, the first two authors agreed on 94.6% of the 74 potential findings; the remaining cases were resolved through discussion.\" · \"…Two authors independently labeled them with 92.1% agreement, and all disagreements were resolved with a third author. This process confirmed 7 false negatives of IAL-SCAN.\"",
     "removal_reason": "",
     "source": "Hou et al. PREPRINT",
     "source_url": "https://arxiv.org/abs/2607.01641",
     "locator": "arXiv:2607.01641v1 §V-B1 and the false-negative paragraph",
     "date_read": "2026-08-14"
    },
    {
     "id": "C101",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C101",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The uncovered inner cycle costs almost nothing at realistic parse-failure rates and only becomes expensive as p climbs. max_iterations=10 throughout, so 10 calls means the bound held",
     "quote": "Run output, python3 Automation/loop_variant_lab.py, 2,000 seeded trials per row: FULL printed table, all fifteen cells ( The lab sweeps p as a chosen parameter and measures nothing about which rate occurs in production.",
     "removal_reason": "",
     "source": "First-party, seeded and deterministic",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C102",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C102",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Call-signature hashing with sorted keys, AND its defeat by semantically-equivalent varied arguments, both shipped with working code",
     "quote": "\"Stable hash of the call signature (sort keys for determinism)\" · \"Sometimes the agent doesn't repeat the exact same call. It does search(\"python async\"), then search(\"async in python\"), then search(\"python asyncio\"). Same intent, different arguments.\" — the article then goes further than we do, to embedding-based semantic detection. Verified at source by the caller. No inference required",
     "removal_reason": "",
     "source": "Alan West, How to Stop Your LLM Agent From Looping Itself Into Oblivion, DEV",
     "source_url": "https://dev.to/alanwest/how-to-stop-your-llm-agent-from-looping-itself-into-oblivion-27eh",
     "locator": "§ the tracker + § varied arguments",
     "date_read": "2026-08-15"
    },
    {
     "id": "C103",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C103",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A runnable no-key simulated runaway predates ours by five months",
     "quote": "\"The LLM calls in this demo are simulated. No API key is required. The budget enforcement is real.\" · \"the agent loops for 30 seconds — ~595 calls, ~$5.95 — before a safety timeout kills it.\" Verified at source by the caller. Its structure is a quality-threshold loop rather than an uncovered retry path (",
     "removal_reason": "",
     "source": "Albert Mavashev, Stop a Runaway AI Agent in Three Lines, runcycles.io, 2026-03-26",
     "source_url": "https://runcycles.io/blog/runaway-demo-agent-cost-blowup-walkthrough",
     "locator": "§ the unguarded run",
     "date_read": "2026-08-15"
    },
    {
     "id": "C104",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C104",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The article's own guard.py prints c740a98e; C22 records a099dc8c, which is loop_lab.py's different call shape.",
     "quote": "Run output from the article's printed listing: repeater -> {'stopped': 'no-progress', 'calls': 3, 'repeated': 'c740a98e'}. Stable across CPython 3.9.6 and 3.12.12 (sha1(json.dumps(call, sort_keys=True)) is deterministic and default separators have not moved in 3.x), so it will match on the reader's machine",
     "removal_reason": "",
     "source": "First-party run of the printed listing",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C105",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C105",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Its counter is on calls, not on time, and neither tools... nor model(messages) carries a timeout",
     "quote": "Executed by the caller with a tool that never returns: hanging tool -> STILL RUNNING after 3s: bounded_run did NOT terminate (SIGALRM at 3s).",
     "removal_reason": "",
     "source": "First-party execution",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C106",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C106",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The OpenAI Agents SDK advances its turn counter and tests the bound before invoking the model",
     "quote": "Read out of the current wheel by the caller: openai-agents 0.21.0, agents/run.py line 1296 current_turn += 1 immediately followed by line 1297 if max_turns is not None and current_turn > max_turns:. Both precede the model invocation",
     "removal_reason": "",
     "source": "First-party, read from the package",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C107",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C107",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Ten successful parses at independent parse-failure rate p cost 10/(1-p) model calls on average, so the simulation corroborates a general law rather than standing on five chosen rates",
     "quote": "Derivation: calls-per-successful-parse is geometric with mean 1/(1-p); ten of them is negative-binomial with mean 10/(1-p). Checked against an independent 20,000-trial simulation written for this row (different generator and seed from loop_variant_lab.py): means 10.2 / 11.1 / 20.0 / 99.9 / 1001.1 against closed-form 10.2 / 11.1 / 20.0 / 100.0 / 1000.0, and medians 10 / 11 / 20 / 97 / 970 reproducing the article's published 10 / 11 / 20 / 97 / 969. The table the article already printed is confirmed by a route that does not use the article's own lab",
     "removal_reason": "",
     "source": "First-party derivation, independently re-simulated",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C109",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C109",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Total calls to reach ten successful parses at failure rate p is negative binomial with r=10 and q=1-p; the article's two columns are that distribution's exact median and 95th percentile, taken from the CDF",
     "quote": "Computed from the exact PMF C(n-1, r-1) q^r p^(n-r) accumulated to the quantile. p=0.02 → median 10, p95 11 · p=0.10 → 11, 13 · p=0.50 → 19, 28 · p=0.90 → 97, 154 · p=0.99 → 967, 1568. C101's 2,000-trial simulation agrees within two counts on every row (its medians 10/11/20/97/969, p95 11/13/28/154/1569), which converts it from the evidence into corroboration by a second method. A sample maximum is not a bound, and printing one in a piece whose thesis is precision about what bounds a loop invited exactly the misreading the piece exists to correct",
     "removal_reason": "",
     "source": "First-party derivation",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C110",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C110",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Flattening two cycles into one is not necessary for termination. A nested retry loop terminates just as well when the inner cycle tests the shared budget on its own path. The necessary property is that every path reaching another model call spends from the same budget before it gets there, and that the test reading that budget ends the loop",
     "quote": "Executed by the caller: the article's nest preserved, with if calls >= max_calls: return inside the inner while, returns {'exit': 'budget', 'calls': 20} — the same 20 as the flattened bounded_run.",
     "removal_reason": "",
     "source": "First-party execution",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-15"
    },
    {
     "id": "C43",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C43",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "This is not an exotic bug",
     "quote": "",
     "removal_reason": "Unsupported. 47 of 6,549 repos is 0.72%, and C8 reports precision with no recall, so the corpus bounds prevalence in neither direction",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-14"
    },
    {
     "id": "C44",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C44",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Forty lines gets you an agent / the forty lines above",
     "quote": "",
     "removal_reason": "Refuted inside the article. Ball says under 400 (C2); the code block is 13 lines. Appeared in the subtitle, a heading and the body",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-14"
    },
    {
     "id": "C45",
     "url": "https://durabilitycurve.com/claims/the-limit-said-10/#c-C45",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Every tutorial in the genre leaves the same three holes / trim is the default in most codebases / most people fail the audit on the first path",
     "quote": "",
     "removal_reason": "Unmeasured population claims presented as observation. Each asserts a distribution over developers' code or the literature with nothing behind it",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-14"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Guardrail Your Agent Can Reach",
    "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/",
    "essay": {
     "title": "The Guardrail Your Agent Can Reach",
     "url": "https://durabilitycurve.com/blog/the-guardrail-your-agent-can-reach/",
     "published": "2026-08-08"
    },
    "last_checked": "2026-08-07",
    "format": "current",
    "counts": {
     "claims": 20,
     "by_status": {
      "VERIFIED": 20
     },
     "verified_at_primary": 20,
     "removed": 2
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/"
   },
   "rows": [
    {
     "id": "L1",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Rule-file gains are largely content-independent — random, shuffled and mismatched-domain files all match curated ones",
     "quote": "\"Performance gains are largely content-independent: random, shuffled, mismatched-domain, and unconverted-format rule files all match curated rules, pointing to a context priming mechanism.\"",
     "removal_reason": "",
     "source": "Zhang et al., Guardrails Beat Guidance",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "Abstract, v2",
     "date_read": "2026-08-07"
    },
    {
     "id": "L2",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L2",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Random rules match expert-curated ones",
     "quote": "",
     "removal_reason": "\"Same effect size\" reads as two equal measured effects; the paper's result is a NULL (Cochran's Q = 4.70, p = 0.697) and the +13.8pp does not itself reach significance against the no-rule baseline (McNemar p = 0.077).",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-07"
    },
    {
     "id": "L3",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Instructions in context files ARE well followed — the strongest evidence against this piece",
     "quote": "\"while instructions in the context files are well followed by coding agents, repository overviews, although popular and recommended by model providers, are not helpful\"",
     "removal_reason": "",
     "source": "Gloaguen et al.",
     "source_url": "https://arxiv.org/abs/2602.11988",
     "locator": "Abstract, v2 (23 Jun 2026)",
     "date_read": "2026-08-07"
    },
    {
     "id": "L4",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Context files do not generally improve task success, and cost over 20% more",
     "quote": "\"Surprisingly, we find that providing context files does not generally improve task success rates, while increasing inference cost by over 20% on average. This observation holds across different LLMs, coding agents, and for both LLM-generated and developer-committed context files.\"",
     "removal_reason": "",
     "source": "Gloaguen et al.",
     "source_url": "https://arxiv.org/abs/2602.11988",
     "locator": "Abstract, v2",
     "date_read": "2026-08-07"
    },
    {
     "id": "L5",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Gloaguen's own conclusion is conditional, not dismissive",
     "quote": "\"We conclude that while context files are useful for specifying non-standard coding practices, any attempts to improve performance should be rigorously evaluated before deployment.\"",
     "removal_reason": "",
     "source": "Gloaguen et al.",
     "source_url": "https://arxiv.org/abs/2602.11988",
     "locator": "Abstract, v2",
     "date_read": "2026-08-07"
    },
    {
     "id": "L6",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Scale of the Zhang study, and of the practice it measures",
     "quote": "\"over 5,000 agent runs of Claude Code with Claude Opus 4.6 on SWE-bench Verified\"; 679 rule files containing 25,532 total rules scraped from GitHub",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "Abstract + method, v2",
     "date_read": "2026-08-07"
    },
    {
     "id": "L7",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Of the 18-rule curated set: 3 shaping (removal hurts), 4 distorting (removal helps), 11 inert",
     "quote": "\"Starting from the full 18-rule curated set (65.7% pass rate), we remove each rule individually and measure the change (Figure 1b). Using a [ABS-DELTA]>5 pp threshold, of 18 rules, 3 are shaping (removal hurts), 4 are distorting (removal helps), and 11 are inert.\" — [ABS-DELTA] renders the paper's absolute-value notation around delta; the pipe characters are substituted because they break a markdown table cell. Nothing else is altered.",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4, Experiment 4a",
     "date_read": "2026-08-07"
    },
    {
     "id": "L8",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Zhang et al. say an already capable coding agent does not need to be told what to do, but benefits from being told what not to do",
     "quote": "\"The base agent already possesses strong coding capabilities; it does not need to be told what to do, but it benefits from being told what not to do.\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4, PBRS reading",
     "date_read": "2026-08-07"
    },
    {
     "id": "L9",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "14 of 18 rules broke at least 2 previously-solved tasks; worst offenders 4 of 17",
     "quote": "\"As shown in Figure 4a, 14 of 18 rules break at least 2 previously-solved tasks; the worst offenders (\"understand full context,\" \"keep functions concise,\" \"do not install dependencies\") each break 4/17 (24%). Only 4 rules are safe (≤1 task broken).\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4, Experiment 4b",
     "date_read": "2026-08-07"
    },
    {
     "id": "L10",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors caveat L9 on run-to-run variance",
     "quote": "\"We treat this evidence as suggestive rather than conclusive: with a baseline retest reliability of 88.2%, each task has an ≈11.8% chance of flipping under run-to-run variance alone, so any single rule's break count cannot be read as fully causal.\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4, Experiment 4b",
     "date_read": "2026-08-07"
    },
    {
     "id": "L11",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Zhang et al. ran a paired comparison on 58 discriminative SWE-bench Verified tasks, and three finer analyses on 35 discriminative tasks with 18 curated rules",
     "quote": "\"We perform three fine-grained analyses on 35 discriminative tasks using a curated set of 18 rules\"; and the study runs \"controlled experiments with a state-of-the-art coding agent on SWE-bench Verified… A paired within-subject design on 58 discriminative tasks\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4 opening; §1",
     "date_read": "2026-08-07"
    },
    {
     "id": "L13",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors name their own selection bias",
     "quote": "\"Selection bias. By selecting discriminative tasks (30–70% baseline pass rate), we enrich for tasks where rules can produce a measurable effect.\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§ Limitations",
     "date_read": "2026-08-07"
    },
    {
     "id": "L14",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Individual harm does not compound when rules are stacked",
     "quote": "\"Finding 5: Individual harm does not compound in ensemble. The robust comparison is between any single rule and the ensemble. If individual rules were genuinely as distortive as 4b suggests, stacking 18 of them should compound…\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§6.4, Finding 5",
     "date_read": "2026-08-07"
    },
    {
     "id": "L15",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "What rules are, in the authors' framing",
     "quote": "Rules \"are persistent (loaded every session), authored by third parties (not the model developer), and intended to shape multi-step tool-using behavior rather than single-turn generation.\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "§1",
     "date_read": "2026-08-07"
    },
    {
     "id": "L16",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic's own position: an instruction is the wrong tool for a must-not-happen",
     "quote": "\"When there's something that absolutely must not happen, an instruction is the wrong tool. Claude will follow the instruction most of the time, but when under pressure, in a long session or an ambiguous situation, or due to a prompt injection in a file accessed as part of the task, the model can fail to follow a prompted rule. A real guardrail needs to be deterministic, and the enforcement methods are hooks and permissions.\"",
     "removal_reason": "",
     "source": "Michael Segner, Anthropic",
     "source_url": "https://claude.com/blog/steering-claude-code-skills-hooks-rules-subagents-and-more",
     "locator": "\"Steering Claude Code\", 18 June 2026",
     "date_read": "2026-08-07"
    },
    {
     "id": "L17",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L17",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "Prompt-only constraint compliance, measured",
     "quote": "",
     "removal_reason": "Benchmark is SWE-bench Lite.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-07"
    },
    {
     "id": "L18",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Zhang et al.'s affiliations — the provenance the piece discloses",
     "quote": "Author footnotes: \"¹AWS Generative AI Innovation Center ²HSBC Holdings Plc., HSBC Technology Center, China\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "Title page, author footnote",
     "date_read": "2026-08-07"
    },
    {
     "id": "L19",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The paper's own headline finding is that polarity matters — i.e. content matters at the level where rules are written",
     "quote": "\"in our data every individually beneficial rule is a negative constraint (\"do not refactor unrelated code\"), while every individually harmful one is a positive directive (\"follow code style\")\"; and the principle: \"constrain what agents must not do, rather than prescribing what they should\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "Abstract, finding (i)",
     "date_read": "2026-08-07"
    },
    {
     "id": "L22",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L22",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "NIST's reference monitor sets three requirements on an enforcement mechanism, and tamperproof is one of them",
     "quote": "\"A set of design requirements on a reference validation mechanism that, as a key component of an operating system, enforces an access control policy over all subjects and objects.\" Three attributes: \"Always invoked (complete mediation)\", \"Tamperproof\", and \"Small enough to be subject to analysis and tests, with verifiable completeness\"",
     "removal_reason": "",
     "source": "NIST, SP 800-53 Rev. 5",
     "source_url": "https://csrc.nist.gov/glossary/term/reference_monitor",
     "locator": "Glossary entry",
     "date_read": "2026-08-07"
    },
    {
     "id": "L21",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic distinguishes deterministic hooks from settings that cannot be overridden locally — the vendor drawing this article's own distinction",
     "quote": "\"Managed settings go further: they are admin-deployed, cannot be overridden by a user's local config, and are the only way to enforce a deterministic, organization-wide guardrail.\" And separately: \"A PreToolUse hook can inspect a call and exit with code 2 to block it.\"",
     "removal_reason": "",
     "source": "Michael Segner, Anthropic",
     "source_url": "https://claude.com/blog/steering-claude-code-skills-hooks-rules-subagents-and-more",
     "locator": "\"Steering Claude Code\", 18 June 2026",
     "date_read": "2026-08-07"
    },
    {
     "id": "L20",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L20",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Rule count does not accumulate damage",
     "quote": "\"Individual rules often appear harmful in isolation yet do not visibly accumulate damage in ensemble: pass rates remain stable across rule counts from 0 to 50.\"",
     "removal_reason": "",
     "source": "Zhang et al.",
     "source_url": "https://arxiv.org/abs/2604.11088",
     "locator": "Abstract, finding (iii)",
     "date_read": "2026-08-07"
    },
    {
     "id": "L12",
     "url": "https://durabilitycurve.com/claims/the-guardrail-your-agent-can-reach/#c-L12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Our own Meadows piece ranks \"system prompts as constraints\" in the same cell as \"tool permissions, action gates\"",
     "quote": "\"Rules of the system (incentives, punishments, constraints)\" / \"Tool permissions, action gates, system prompts as constraints\"",
     "removal_reason": "",
     "source": "The Leverage Hierarchy of Agent Engineering (ours)",
     "source_url": "https://harryfloyd.substack.com/p/the-leverage-hierarchy-of-agent-engineering",
     "locator": "The 12-rank table, rank 5",
     "date_read": "2026-08-07"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Average Is Nobody's Result",
    "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/",
    "essay": {
     "title": "The Average Is Nobody's Result",
     "url": "https://durabilitycurve.com/blog/the-average-is-nobodys-result/",
     "published": "2026-08-04"
    },
    "last_checked": "2026-09-27",
    "format": "current",
    "counts": {
     "claims": 35,
     "by_status": {
      "EXECUTED": 11,
      "VERIFIED": 23,
      "REPORTED": 1
     },
     "verified_at_primary": 23,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-average-is-nobodys-result/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-1",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Of 255 abstracts with text in the corpus, 21 report the assistant's effect split by a property of the operator and 234 report an average and nothing else",
     "quote": "\"HEADLINE: 21/255 studies report the assistant's effect split by a property of the OPERATOR. = 8.2% 234 report an average over operators and nothing else.\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-2",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Among primary records by PubMed's own publication type the figure is 21 of 157",
     "quote": "\"Among primary records only: 21/157 = 13.4%\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-3",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Of the 21, only 8 print per-stratum effect estimates for every stratum; 3 print one stratum and a bare null; 10 report only a direction or a significance verdict",
     "quote": "\"SPLIT_FULL 8 per-stratum estimates for every stratum · SPLIT_PARTIAL 3 one stratum estimated, the other a bare null · SPLIT_DIRECTION 10 a direction or a p-verdict, nothing poolable\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-4",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The splits disagree: 7 find the larger effect in weaker operators, 4 in stronger, 3 no interaction, 3 nothing significant after stratifying, 1 divergence",
     "quote": "\"LOWER 7 ... HIGHER 4 ... UNIFORM 3 ... NONE 3 ... DIVERGES 1\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-5",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Three deliberately dissimilar detectors were run; the obvious one reached 15 records of which 7 are real splits, missing 14",
     "quote": "\"pattern A: reached 15 of which real splits 7 precision 0.47 coverage 0.33 of the 21 found\" · \"Pattern A alone, unread, would have reported 15 records of which 7 are real, missing 14.\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-6",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Pattern B reached 39 records at precision 0.36 and coverage 0.67 of the found set; pattern C reached 13 at precision 0.69 and coverage 0.43",
     "quote": "\"pattern B: reached 39 of which real splits 14 precision 0.36 coverage 0.67 of the 21 found\" · \"pattern C: reached 13 of which real splits 9 precision 0.69 coverage 0.43 of the 21 found\" · \"'coverage' is the share of the found positives a pattern reached, NOT recall\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-7",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Every record any pattern reached was hand-read, and the script refuses to print the headline while one is unread",
     "quote": "\"HEADLINE: UNAVAILABLE — 20 record(s) reached by a pattern are unread.\" (the refusal branch, observed firing before adjudication was complete)",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-8",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "The PubMed query returns 259 records, 255 with abstracts",
     "quote": "\"PubMed hits: 259 (retmax 400, fetched 259)\" · \"classified : 255 (no abstract: 4)\"",
     "removal_reason": "",
     "source": "first-party run by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-9",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "In a seeded 20-record sample of the primary average-only class, 9 had open full text and none reported an operator split in its results",
     "quote": "Sample drawn with random.Random(20260802) from 136 eligible records; PMC full text retrieved for 9; operator-split patterns returned 0 result-section hits in 8, and in the ninth the only hit was future-tense text in a trial protocol",
     "removal_reason": "",
     "source": "first-party full-text check by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-10",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Eleven of the twenty sampled records are not in PMC and their full texts were not read",
     "quote": "NCBI ID converter returned \"Identifier not found in PMC\" for 11 of the 20 sampled PMIDs",
     "removal_reason": "",
     "source": "first-party full-text check by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Reanalysing a CAD mammography study, use of CAD was associated with a 0.016 increase in sensitivity (95% CI 0.003-0.028) for the 44 least discriminating radiologists on 45 relatively easy cancers",
     "quote": "\"Use of CAD was associated with a 0.016 increase in sensitivity (95% confidence interval [CI], 0.003-0.028) for the 44 least discriminating radiologists for 45 relatively easy, mostly CAD-detected cancers.\"",
     "removal_reason": "",
     "source": "Povyakalo, Alberdi, Strigini & Ayton, Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Results",
     "date_read": "2026-08-02"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "For the 6 most discriminating radiologists, sensitivity decreased by 0.145 (95% CI 0.034-0.257) on the 15 relatively difficult cancers",
     "quote": "\"However, for the 6 most discriminating radiologists, with CAD, sensitivity decreased by 0.145 (95% CI, 0.034-0.257) for the 15 relatively difficult cancers.\"",
     "removal_reason": "",
     "source": "Povyakalo et al., Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Results",
     "date_read": "2026-08-02"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The original study detected no significant average effect, and the reanalysis found CAD helped the less discriminating readers and hindered the more discriminating ones",
     "quote": "\"It indicates that, despite the original study detecting no significant average effect, CAD helped the less discriminating readers but hindered the more discriminating readers.\"",
     "removal_reason": "",
     "source": "Povyakalo et al., Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Conclusions",
     "date_read": "2026-08-02"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "That reanalysis covered 50 readers interpreting 180 mammograms both with and without computer support",
     "quote": "\"reanalyzing data from a published study where 50 professionals (\\\"readers\\\") interpreted 180 mammograms, both with and without computer support\"",
     "removal_reason": "",
     "source": "Povyakalo et al., Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Background",
     "date_read": "2026-08-02"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The reader and case strata were derived after the fact from the same regression, and the authors call the method exploratory",
     "quote": "\"Using regression estimates, we obtained thresholds for classifying a posteriori the cases (by difficulty) and the readers (by discriminating ability).\" · \"Our exploratory analysis method reveals unexpected effects.\"",
     "removal_reason": "",
     "source": "Povyakalo et al., Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Method and Conclusions",
     "date_read": "2026-08-02"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The authors said such differential effects should be assessed when evaluating CAD and similar warning systems",
     "quote": "\"Such differential effects, although subtle, may be clinically significant and important for improving both computer algorithms and protocols for their use. They should be assessed when evaluating CAD and similar warning systems.\"",
     "removal_reason": "",
     "source": "Povyakalo et al., Med Decis Making 2013",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/23300205/",
     "locator": "deposited abstract, Conclusions",
     "date_read": "2026-08-02"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a population-based randomised trial of 4,824 surveillance colonoscopies, CADe increased ADR among low-performing endoscopists (45.5% vs 52.1%; aRR 1.15, 95% CI 1.01-1.30) but not among high-performing ones (65.9% vs 63.5%; aRR 0.96, 95% CI 0.88-1.05)",
     "quote": "\"Of 5,309 randomized surveillance colonoscopies, 4,824 surveillance colonoscopies were included in the final analysis\" · \"CADe increased ADR among low-performing (ADR<54.5%) endoscopists (45.5% vs 52.1%; aRR 1.15 [95% CI 1.01-1.30]), but not among high-performing endoscopists (65.9% vs 63.5%; aRR 0.96 [95% CI 0.88-1.05]).\"",
     "removal_reason": "",
     "source": "Endoscopy 2026, Galician screening programme, PMID 42365851",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/42365851/",
     "locator": "deposited abstract, Results",
     "date_read": "2026-08-02"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "That subgroup analysis was prespecified",
     "quote": "\"Prespecified subgroup analyses evaluated endoscopist baseline performance (median ADR in the standard arm).\"",
     "removal_reason": "",
     "source": "Endoscopy 2026, PMID 42365851",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/42365851/",
     "locator": "deposited abstract, Methods",
     "date_read": "2026-08-02"
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a 3,059-patient multicentre randomised trial the gain was larger in experts: ADR of experts 42.3% vs 32.8% (P < .001) and of nonexperts 37.5% vs 32.1% (P = .023)",
     "quote": "\"From November 2019 to August 2021, 3059 subjects were randomized to AI-assisted colonoscopy (n = 1519) and conventional colonoscopy (n = 1540).\" · \"ADR of expert (42.3% vs 32.8%; P < .001) and nonexpert endoscopists (37.5% vs 32.1%; P = .023)... were all significantly higher in the AI-assisted colonoscopy.\"",
     "removal_reason": "",
     "source": "Clin Gastroenterol Hepatol 2023, PMID 35863686",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/35863686/",
     "locator": "deposited abstract, Results",
     "date_read": "2026-08-02"
    },
    {
     "id": "20",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-20",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A 2026 study expected the greatest increase among low-volume and junior endoscopists",
     "quote": "\"The greatest increase was expected among low-volume and junior endoscopists.\"",
     "removal_reason": "",
     "source": "Dis Colon Rectum 2026, PMID 41919624",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/41919624/",
     "locator": "deposited abstract, Objective",
     "date_read": "2026-08-02"
    },
    {
     "id": "21",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "It reported that seniors rose from 51.7% to 59.6% (p = 0.02) and juniors from 56.2% to 62.9% (p = 0.12), and concluded the tool correlated with increased detection for experienced endoscopists",
     "quote": "\"Of 12 senior and 12 junior endoscopists, the seniors had a statistically significant increase ( p = 0.02) from 51.7% to 59.6%, whereas juniors did not (56.2% to 62.9%, p = 0.12).\" · \"Computer-aided detection in colonoscopy correlated with an increased adenoma detection rate for experienced endoscopists.\"",
     "removal_reason": "",
     "source": "Dis Colon Rectum 2026, PMID 41919624",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/41919624/",
     "locator": "deposited abstract, Results and Conclusions",
     "date_read": "2026-08-02"
    },
    {
     "id": "22",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-22",
     "status": "EXECUTED",
     "status_label": "Executed",
     "removed": false,
     "claim": "Those two reported increases are 7.9 points and 6.7 points, a gap of 1.2 points, and only the larger crossed p<0.05",
     "quote": "59.6 − 51.7 = 7.9 · 62.9 − 56.2 = 6.7 · 7.9 − 6.7 = 1.2. Subtraction only; no model, no reanalysis, and no claim that either estimate is wrong",
     "removal_reason": "",
     "source": "first-party arithmetic on row 21's quoted rates by The Durability Curve",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "23",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-23",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same study reported the same pattern on two further operator axes: high-volume 51.3% to 59.4% (p = 0.01) against low-volume 56.5% to 63.1% (p = 0.13), and surgeons 49% to 62% (p = 0.02) against gastroenterologists 56.9% to 60.8% (p = 0.15)",
     "quote": "\"Colorectal surgeons had a statistically significant increase in adenoma detection rate from 49% to 62% ( p = 0.02), but gastroenterologists did not (56.9% to 60.8%, p = 0.15).\" · \"When comparing 12 high-volume and 12 low-volume endoscopists, the high-volume group had a statistically significant increase (51.3% to 59.4%, p = 0.01), whereas the low-volume group did not (56.5% to 63.1%, p = 0.13).\"",
     "removal_reason": "",
     "source": "Dis Colon Rectum 2026, PMID 41919624",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/41919624/",
     "locator": "deposited abstract, Results",
     "date_read": "2026-08-02"
    },
    {
     "id": "24",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-24",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Pooling two randomised trials, CADe raised ADR (RR 1.29, 95% CI 1.16 to 1.42) while examiner experience did not (RR 1.02, 95% CI 0.89 to 1.16), and the authors concluded experience plays a minor role",
     "quote": "\"use of CADe (RR 1.29; 95% CI: 1.16 to 1.42) and colonoscopy indication, but not the level of examiner experience (RR 1.02; 95% CI: 0.89 to 1.16) were associated with ADR differences in a multivariate analysis.\" · \"Experience appears to play a minor role as determining factor for ADR.\"",
     "removal_reason": "",
     "source": "Gut 2022, AID-1 + AID-2, PMID 34187845",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/34187845/",
     "locator": "deposited abstract, Results and Conclusion",
     "date_read": "2026-08-02"
    },
    {
     "id": "25",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-25",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The title of the paper naming the inference error is \"The Difference Between 'Significant' and 'Not Significant' is not Itself Statistically Significant\"",
     "quote": "\"The Difference Between \\\"Significant\\\" and \\\"Not Significant\\\" is not Itself Statistically Significant\"",
     "removal_reason": "",
     "source": "Gelman & Stern, The American Statistician 60(4):328-331, 2006",
     "source_url": "http://www.stat.columbia.edu/~gelman/research/published/signif4.pdf",
     "locator": "author's deposited PDF, title, page 1",
     "date_read": "2026-08-02"
    },
    {
     "id": "26",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-26",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Their point is not that thresholds are arbitrary but that large changes in significance can correspond to small, non-significant changes in the underlying quantities",
     "quote": "\"Rather, we are pointing out that even large changes in significance levels can correspond to small, nonsignificant changes in the underlying quantities.\"",
     "removal_reason": "",
     "source": "Gelman & Stern, 2006",
     "source_url": "http://www.stat.columbia.edu/~gelman/research/published/signif4.pdf",
     "locator": "deposited PDF, abstract, page 1",
     "date_read": "2026-08-02"
    },
    {
     "id": "27",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-27",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The reporting standard for AI trials asks investigators to state what level of expertise was required of users",
     "quote": "\"CONSORT-AI 5 (iv) Extension Specify whether there was human–AI interaction in the handling of the input data, and what level of expertise was required of users.\"",
     "removal_reason": "",
     "source": "CONSORT-AI extension, Nat Med 26:1364-1374, 2020",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7598943/",
     "locator": "checklist table, item 5 (iv), PMC full text",
     "date_read": "2026-08-02"
    },
    {
     "id": "28",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-28",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Its guidance encourages exploring differences in performance across population subgroups, and no item asks for the result across operator subgroups",
     "quote": "\"Beyond this, investigators should also be encouraged to explore differences in performance and error rates across population subgroups.\"",
     "removal_reason": "",
     "source": "CONSORT-AI extension, 2020",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7598943/",
     "locator": "discussion of item 19 extension, PMC full text; the full 14-item extension list was read and contains no operator-stratified reporting item",
     "date_read": "2026-08-02"
    },
    {
     "id": "29",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-29",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The COLO-DETECT protocol pre-registered a subgroup analysis by colonoscopist type",
     "quote": "\"Subgroup analyses will be conducted on colonoscopist type (i.e., non-BCSP accredited vs. BCSP accredited) and indication for colonoscopy (screening vs. symptomatic).\"",
     "removal_reason": "",
     "source": "COLO-DETECT trial protocol, Colorectal Dis 2022, PMID 35680613",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC9796278/",
     "locator": "protocol full text, statistical analysis section, PMC9796278",
     "date_read": "2026-08-02"
    },
    {
     "id": "30",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-30",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same protocol said a range of colonoscopist experience was anticipated and desirable, and specified how it would be measured",
     "quote": "\"A range of experience amongst participating colonoscopists is anticipated and desirable; to facilitate subgroup analysis by colonoscopist experience, it will be assessed by BCSP accreditation status, lifetime procedure numbers and yearly procedure numbers for the last 3 years\"",
     "removal_reason": "",
     "source": "COLO-DETECT protocol, 2022",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC9796278/",
     "locator": "protocol full text, PMC9796278",
     "date_read": "2026-08-02"
    },
    {
     "id": "31",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-31",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The COLO-DETECT results reported one average across 2032 participants at 12 NHS hospitals: adenomas detected in 56·6% versus 48·4%, adjusted odds ratio 1·47 (95% CI 1·21-1·78)",
     "quote": "\"We did a multicentre, open-label, parallel-arm, pragmatic randomised controlled trial in 12 National Health Service (NHS) hospitals (ten NHS Trusts) in England\" · \"Between March 29, 2021, and April 6, 2023, 2032 participants (1132 [55·7%] male, 900 [44·3%] female; mean age 62·4 years [SD 10·8]) were recruited and randomly assigned\" · \"Adenomas were detected in 555 (56·6%) of 980 participants in the CADe-assisted colonoscopy group versus 477 (48·4%) of 986 in the standard colonoscopy group, representing a proportion difference of 8·3% (95% CI 3·9-12·7; adjusted odds ratio 1·47 [95% CI 1·21-1·78], p<0·0001).\"",
     "removal_reason": "",
     "source": "COLO-DETECT, Lancet Gastroenterol Hepatol 2024, PMID 39153491",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/39153491/",
     "locator": "deposited abstract, Methods and Findings",
     "date_read": "2026-08-02"
    },
    {
     "id": "32",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-32",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Its randomisation was stratified by age group, sex, indication and NHS Trust, and its abstract names no colonoscopist subgroup",
     "quote": "Stratification verbatim: \"with stratification by age group, sex, colonoscopy indication (screening or symptomatic), and NHS Trust\". The absence of a colonoscopist subgroup is an absence in the deposited abstract only — the paper is not in PMC and the full text was not read.",
     "removal_reason": "",
     "source": "COLO-DETECT, Lancet Gastro Hep 2024, PMID 39153491",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/39153491/",
     "locator": "deposited abstract, Methods",
     "date_read": "2026-08-02"
    },
    {
     "id": "33",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-33",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The field's current answer is a study-level one: pooling twenty-eight randomised trials and 23,861 participants, an expert-only subgroup showed a similar effect size (RR 1.19, 95% CI 1.11-1.27) and the review concluded the benefit holds irrespective of endoscopist experience",
     "quote": "\"Twenty-eight RCTs were selected for inclusion involving 23,861 participants.\" · \"Subgroup analyses involving only expert endoscopists demonstrated a similar effect size (RR, 1.19; 95% CI, 1.11-1.27; P < .001)\" · \"AI-assisted colonoscopy significantly improved adenoma detection but not sessile serrated lesion detection irrespective of endoscopist experience, system type, or healthcare setting.\"",
     "removal_reason": "",
     "source": "Gastrointest Endosc 2025 systematic review, PMID 39216648",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/39216648/",
     "locator": "deposited abstract, Results and Conclusions",
     "date_read": "2026-09-27"
    },
    {
     "id": "34",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-34",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A second meta-analysis reached the same study-level conclusion across twenty-four randomised trials and 17,413 colonoscopies",
     "quote": "\"Twenty-four RCTs involving 17,413 colonoscopies (AI assisted: 8680; non-AI assisted: 8733) were included.\" · \"Type of AI system used or endoscopist experience did not affect overall improvement in ADR.\"",
     "removal_reason": "",
     "source": "Gastrointest Endosc 2024, PMID 38272274",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/38272274/",
     "locator": "deposited abstract, Results and Conclusions",
     "date_read": "2026-08-02"
    },
    {
     "id": "35",
     "url": "https://durabilitycurve.com/claims/the-average-is-nobodys-result/#c-35",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Bainbridge stated in 1983 that unused physical skills decay and a monitoring operator becomes an inexperienced one",
     "quote": "\"Unfortunately, physical skills deteriorate when they are not used, particularly the refinements of gain and timing. This means that a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one.\"",
     "removal_reason": "",
     "source": "Bainbridge, Automatica 19(6):775-779, 1983",
     "source_url": "https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf",
     "locator": "§1.1.1 Manual control skills",
     "date_read": "2026-08-01"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: You Cannot Try to Fall Asleep",
    "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/",
    "essay": {
     "title": "You Cannot Try to Fall Asleep",
     "url": "https://durabilitycurve.com/blog/you-cannot-try-to-fall-asleep/",
     "published": "2026-08-03"
    },
    "last_checked": "2026-08-02",
    "format": "current",
    "counts": {
     "claims": 20,
     "by_status": {
      "VERIFIED": 15,
      "REPORTED": 5
     },
     "verified_at_primary": 15,
     "removed": 1
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Sleep normalcy is a relatively automatic process. Consequently it is vulnerable, and may be inhibited, by focused attention and by direct attempts to control its expression",
     "quote": "\"The argument is that sleep normalcy is a relatively automatic process. Consequently, it is vulnerable, and may be inhibited, by focused attention and by direct attempts to control its expression.\"",
     "removal_reason": "",
     "source": "Espie, Broomfield, MacMahon, Macphee & Taylor, The attention-intention-effort pathway in the development of psychophysiologic insomnia: a theoretical review, Sleep Medicine Reviews 10(4):215-45, Aug 2006, PMID 16809056, DOI 10.1016/j.smrv.2006.03.002",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/16809056/",
     "locator": "abstract, sentences 6-7",
     "date_read": "2026-08-02"
    },
    {
     "id": "1b",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-1b",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The model is named the attention-intention-effort pathway, and it is offered as an explanatory model of how psychophysiologic insomnia develops and is maintained",
     "quote": "\"This paper proposes an explanatory model, that we call the attention-intention-effort pathway.\"",
     "removal_reason": "",
     "source": "Espie et al. 2006, as above",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/16809056/",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "1c",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-1c",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Psychophysiologic insomnia is the most common form of persistent primary insomnia, and its behavioural phenotype includes conditioned arousal, sleep-incompatible behaviour and sleep preoccupation",
     "quote": "\"Psychophysiologic insomnia (PI) is the most common form of persistent primary insomnia. Its 'behavioral phenotype', comprising elements such as conditioned arousal, sleep-incompatible behavior and sleep preoccupation, has not changed markedly across several generations of diagnostic nosology.\"",
     "removal_reason": "",
     "source": "Espie et al. 2006, as above",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/16809056/",
     "locator": "abstract, opening",
     "date_read": "2026-08-02"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-2",
     "status": "STRUCK",
     "status_label": "Struck",
     "removed": true,
     "claim": "The three components act in concert and hierarchically to turn acute stress-induced insomnia into the persistent self-perpetuating kind",
     "quote": "",
     "removal_reason": "This wording is not in the 2006 abstract; it came from a secondary summary of the 2023 revisit, which is paywalled (HTTP 402) and was not read. The body uses rows 1/1b/1c instead and makes no claim about hierarchy or about the acute-to-chronic transition",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": "2026-08-02"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Twenty-two individuals with primary insomnia were given positive or negative feedback about their sleep immediately on waking on three consecutive mornings",
     "quote": "\"twenty-two individuals with primary insomnia receiv[ed] positive or negative feedback about their sleep immediately on waking on three consecutive mornings\"",
     "removal_reason": "",
     "source": "Semler & Harvey, Misperception of sleep can adversely affect daytime functioning in insomnia, Behaviour Research and Therapy 43(7):843–856, 2005, PMID 15896282",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/15896282/",
     "locator": "method",
     "date_read": "2026-08-02"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Objective sleep, by actigraphy, did not differ across the three nights or the two feedback conditions",
     "quote": "\"Objective sleep on each of the three nights was estimated by actigraphy and did not differ across the three nights or the two feedback conditions\"",
     "removal_reason": "",
     "source": "Semler & Harvey 2005",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/15896282/",
     "locator": "results",
     "date_read": "2026-08-02"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Negative feedback was associated with more negative thoughts, sleepiness, monitoring for sleep-related threat, and safety behaviours during the day",
     "quote": "\"negative feedback was associated with more negative thoughts, sleepiness, monitoring for sleep-related threat, and safety behaviours during the day, relative to positive feedback\"",
     "removal_reason": "",
     "source": "Semler & Harvey 2005",
     "source_url": "https://pubmed.ncbi.nlm.nih.gov/15896282/",
     "locator": "results",
     "date_read": "2026-08-02"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Fifty-four healthy sleepers were randomly allocated to receive good or poor false sleep feedback in the form of a numerical sleep score, and were told it reflected their habitual sleep",
     "quote": "\"A total of 54 healthy sleepers (Mage = 30.19 years; SDage = 12.94 years) were randomly allocated to receive good, or poor, false sleep feedback, in the form of a numerical sleep score. Participants were informed that this feedback was a true reflection of their habitual sleep.\"",
     "removal_reason": "",
     "source": "Robson, Ellis & Elder, Poor false sleep feedback does not affect pre-sleep cognitive arousal or subjective sleep continuity in healthy sleepers: a pilot study, Sleep and Biological Rhythms 20(4):467–472, 2022",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10899903/",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In healthy sleepers the false score changed nothing measured",
     "quote": "\"There were no significant differences between good and poor feedback groups in terms of pre-sleep cognitive arousal, or subjective sleep continuity, before or after the presentation of the sleep feedback.\"",
     "removal_reason": "",
     "source": "Robson, Ellis & Elder 2022",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10899903/",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "That study frames itself as being about wearables, not about laboratory deception",
     "quote": "\"Modern wearable devices calculate a numerical metric of sleep quality (sleep feedback), which are intended to allow users to monitor and, potentially, improve their sleep.\"",
     "removal_reason": "",
     "source": "Robson, Ellis & Elder 2022",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10899903/",
     "locator": "abstract, first line",
     "date_read": "2026-08-02"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Measurement increases how much of an activity people do and simultaneously reduces how much they enjoy it",
     "quote": "\"while measurement increases how much of an activity people do (e.g., walk or read more), it can simultaneously reduce how much people enjoy those activities\"",
     "removal_reason": "",
     "source": "Etkin, The Hidden Cost of Personal Quantification, Journal of Consumer Research 42(6):967–984, 2016",
     "source_url": "https://academic.oup.com/jcr/article-abstract/42/6/967/2358309",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The mechanism is that attention to output makes an enjoyable activity feel like work, undermining intrinsic motivation",
     "quote": "\"By drawing attention to output, measurement can make enjoyable activities feel more like work, which reduces their enjoyment.\"",
     "removal_reason": "",
     "source": "Etkin 2016",
     "source_url": "https://academic.oup.com/jcr/article-abstract/42/6/967/2358309",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Etkin's 2016 findings about measuring an activity rest on six experiments",
     "quote": "\"Six experiments demonstrate that while measurement increases how much of an activity people do…\"",
     "removal_reason": "",
     "source": "Etkin 2016",
     "source_url": "https://academic.oup.com/jcr/article-abstract/42/6/967/2358309",
     "locator": "abstract",
     "date_read": "2026-08-02"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A state is essentially a by-product when it can come about as a result of action but cannot be brought about intentionally, because the attempt precludes the state",
     "quote": "\"can only come about as by-products of actions undertaken for other ends… can never be brought about intentionally because the very attempt to do so precludes the state one is trying to bring about\"",
     "removal_reason": "",
     "source": "Elster, Sour Grapes: Studies in the Subversion of Rationality, Cambridge University Press, ch. 2 \"States that are essentially by-products\", pp. 43–109",
     "source_url": "https://doi.org/10.1017/CBO9781316494172.004",
     "locator": "chapter 2 summary",
     "date_read": "2026-08-02"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Elster names spontaneity as an example of such an inaccessible state",
     "quote": "\"I cited spontaneity as an example of such 'inaccessible' states.\"",
     "removal_reason": "",
     "source": "Elster, Sour Grapes ch. 2",
     "source_url": "https://doi.org/10.1017/CBO9781316494172.004",
     "locator": "chapter 2 introduction",
     "date_read": "2026-08-02"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-15",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "The self-defeating list includes the desire to forget, to believe, to desire, to sleep, to laugh, and to overcome stuttering",
     "quote": "\"the desire to forget, the desire to believe, the desire to desire …, the desire to sleep, the desire to laugh (one cannot tickle oneself), and the desire to overcome stuttering\"",
     "removal_reason": "",
     "source": "Elster, Sour Grapes ch. 2 / States that are essentially by-products, Social Science Information 20(3), 1981",
     "source_url": "https://journals.sagepub.com/doi/10.1177/053901848102000301",
     "locator": "ch. 2, as reproduced in secondary summaries",
     "date_read": "2026-08-02"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Chapter 2 of Sour Grapes is titled \"States that are essentially by-products\" and runs pp. 43-109",
     "quote": "\"Chapter: 2 ('States that are essentially by-products') · Pages: 43-109\"",
     "removal_reason": "",
     "source": "Cambridge Core chapter listing",
     "source_url": "https://doi.org/10.1017/CBO9781316494172.004",
     "locator": "publisher metadata block",
     "date_read": "2026-08-02"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-17",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Orthosomnia names patients whose pursuit of the sleep score interfered with their sleep, and who trusted the device over the clinician",
     "quote": "",
     "removal_reason": "",
     "source": "Baron, Abbott, Jao, Manalo & Mullen, Orthosomnia: Are Some Patients Taking the Quantified Self Too Far?, Journal of Clinical Sleep Medicine, 2017",
     "source_url": "https://jcsm.aasm.org/doi/10.5664/jcsm.6472",
     "locator": "—",
     "date_read": "2026-08-02"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-18",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Whoop's Recovery Score is delivered each morning and is computed from HRV measured during slow-wave sleep, resting heart rate, respiratory rate and sleep performance",
     "quote": "\"The Recovery Score (0-100) … a daily readiness indicator that aggregates heart rate variability, resting heart rate and sleep quality … reported each morning after your sleep cycle completes\"",
     "removal_reason": "",
     "source": "Whoop product documentation",
     "source_url": "https://developer.whoop.com/docs/whoop-101/",
     "locator": "vendor marketing and developer documentation",
     "date_read": "2026-08-02"
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-19",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "A 2024 cross-sectional study of 523 adults estimated orthosomnia prevalence at 3–14% depending on definition, with 35.8% of participants regularly using sleep-tracking devices",
     "quote": "— figures from secondary summary; primary not yet opened —",
     "removal_reason": "",
     "source": "Brain Sciences, 2024",
     "source_url": "https://doaj.org/article/50f708de43bd4086a44688d71f87f333",
     "locator": "—",
     "date_read": "2026-08-02"
    },
    {
     "id": "20",
     "url": "https://durabilitycurve.com/claims/you-cannot-try-to-fall-asleep/#c-20",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "A validated Bergen Orthosomnia Scale was published in 2025",
     "quote": "",
     "removal_reason": "",
     "source": "secondary summary only",
     "source_url": "",
     "locator": "—",
     "date_read": "2026-08-02"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Your Robot Coworker Is Still a Pilot",
    "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/",
    "essay": {
     "title": "Your Robot Coworker Is Still a Pilot",
     "url": "https://durabilitycurve.com/blog/your-robot-coworker-is-still-a-pilot/",
     "published": "2026-08-01"
    },
    "last_checked": "2026-08-01",
    "format": "current",
    "counts": {
     "claims": 24,
     "by_status": {
      "VERIFIED": 22,
      "REPORTED": 2
     },
     "verified_at_primary": 22,
     "removed": 5
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/"
   },
   "rows": [
    {
     "id": "1",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Unitree published a clarification because its 2025 shipment volume was being misreported",
     "quote": "\"many pieces of misinformation regarding our company's 2025 shipment volume have been circulating online\"",
     "removal_reason": "",
     "source": "Unitree Robotics, official news post",
     "source_url": "https://shop.unitree.com/blogs/news/clarification-regarding-unitrees-2025-sales-data",
     "locator": "opening paragraph",
     "date_read": "2026-08-01"
    },
    {
     "id": "2",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Unitree shipped more than 5,500 humanoid robots in 2025",
     "quote": "\"Unitree's actual shipment volume of humanoid robots exceeded 5,500 units\"",
     "removal_reason": "",
     "source": "Unitree Robotics, 2025 sales clarification, 22 Jan 2026",
     "source_url": "https://shop.unitree.com/blogs/news/clarification-regarding-unitrees-2025-sales-data",
     "locator": "shipment figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "3",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Unitree produced more than 6,500 humanoid robots in 2025, a different and larger number than it shipped",
     "quote": "\"total mass-production output of 2025 exceeded 6,500 units\"",
     "removal_reason": "",
     "source": "Unitree Robotics, 2025 sales clarification, 22 Jan 2026",
     "source_url": "https://shop.unitree.com/blogs/news/clarification-regarding-unitrees-2025-sales-data",
     "locator": "production figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "4",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Unitree states its shipment figure counts delivery to end customers, and that its order volume is higher still",
     "quote": "\"quantity actually sold and delivered to end customers, not order volume; the order volume is higher\"",
     "removal_reason": "",
     "source": "Unitree Robotics, 2025 sales clarification, 22 Jan 2026",
     "source_url": "https://shop.unitree.com/blogs/news/clarification-regarding-unitrees-2025-sales-data",
     "locator": "definition note",
     "date_read": "2026-08-01"
    },
    {
     "id": "5",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Unitree warns against adding different robot types together",
     "quote": "\"directly combining the numbers of different types of robots together for comparison\"",
     "removal_reason": "",
     "source": "Unitree Robotics, 2025 sales clarification, 22 Jan 2026",
     "source_url": "https://shop.unitree.com/blogs/news/clarification-regarding-unitrees-2025-sales-data",
     "locator": "closing caution",
     "date_read": "2026-08-01"
    },
    {
     "id": "6",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility Robotics announced a 2.5 billion dollar merger with Churchill Capital Corp XI",
     "quote": "\"Agility Robotics to Go Public Through $2.5 Billion Merger with Churchill Capital Corp XI\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.1, joint press release, 24 June 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-1.htm",
     "locator": "headline",
     "date_read": "2026-08-01"
    },
    {
     "id": "7",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility reports 300 million dollars of contracted orders rather than a number of robots",
     "quote": "\"more than $300 million of multi-year contracted Digit v5 orders secured to date\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.1, 24 June 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-1.htm",
     "locator": "highlights list",
     "date_read": "2026-08-01"
    },
    {
     "id": "8",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility reports 9 customer facilities and 65,000 operating hours in place of a robot count",
     "quote": "\"Through deployment commitments across nine customer facilities, Digit has accumulated more than 65,000 hours of operation\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.1, 24 June 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-1.htm",
     "locator": "commercial traction paragraph",
     "date_read": "2026-08-01"
    },
    {
     "id": "9",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility's 10,000 unit figure describes annual factory capacity, not robots made",
     "quote": "\"RoboFab, its full-scale humanoid manufacturing facility designed to support production of up to 10,000 units annually\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.1, 24 June 2026",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-1.htm",
     "locator": "manufacturing paragraph",
     "date_read": "2026-08-01"
    },
    {
     "id": "10",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility's press-release exhibit carries no robot unit count; the count sits in the companion investor-presentation exhibit of the same filing",
     "quote": "\"more than $300 million of multi-year contracted Digit v5 orders secured to date\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.1 (press release), 24 June 2026; parsed in full, no unit count present",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-1.htm",
     "locator": "whole press release, six pages",
     "date_read": "2026-08-01"
    },
    {
     "id": "11",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility's investor presentation discloses one robot count, 1,000, and defines it as customer ORDERS not built/shipped/installed units",
     "quote": "\"Reflects customer orders for Digit v5, as of May 2026 … and relates to 1,000 Digit v5 robots with three-year term RaaS contract\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.2 (Investor Presentation, June 2026), same 8-K accession 000121390026071290",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-2.htm",
     "locator": "footnote (1) to the orders figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "12",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same RaaS contract ties buyer warrants to robots DEPLOYED, treating deployment as a separate event from the order and giving it no number",
     "quote": "\"includes warrants issued to purchaser vesting proportionately to robots deployed; figures are not a measure of current period revenue\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.2 (Investor Presentation, June 2026)",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-2.htm",
     "locator": "footnote (1) to the orders figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "13",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility's presentation says its robots are currently deployed in customer facilities but gives no deployed count; its unit-scale numbers are production scenarios and an illustrative chart",
     "quote": "\"currently deployed in customer facilities\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.2 (Investor Presentation, June 2026)",
     "source_url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071290/ea029548401ex99-2.htm",
     "locator": "metrics slide; the \"1k/5k/10k units/yr\" cost curve and the \"installed base\" revenue chart are marked \"for illustrative purposes only\"",
     "date_read": "2026-08-01"
    },
    {
     "id": "14",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "BMW says the Spartanburg robot moved more than 90,000 components over about 1,250 operating hours",
     "quote": "\"In total, it moved more than 90,000 components and covered approximately 1.2 million steps in around 1,250 operating hours\"",
     "removal_reason": "",
     "source": "BMW Group press release, 27 Feb 2026",
     "source_url": "https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time?language=en",
     "locator": "Spartanburg results",
     "date_read": "2026-08-01"
    },
    {
     "id": "15",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-15",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "BMW says that within ten months the robot supported production of more than 30,000 X3 vehicles (secondary coverage's \"eleven months\" is not BMW's own figure)",
     "quote": "\"Within ten months, the robot Figure 02 supported the production of more than 30,000 BMW X3, working ten-hour shifts daily from Monday to Friday\"",
     "removal_reason": "",
     "source": "BMW Group press release, 27 Feb 2026",
     "source_url": "https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time?language=en",
     "locator": "Spartanburg results",
     "date_read": "2026-08-01"
    },
    {
     "id": "16",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-16",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "BMW's release refers to the robot in the singular and never states how many robots were at Spartanburg",
     "quote": "\"the robot Figure 02 supported the production of more than 30,000 BMW X3\"",
     "removal_reason": "",
     "source": "BMW Group press release, 27 Feb 2026; whole release checked for a unit count, none present",
     "source_url": "https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time?language=en",
     "locator": "whole release",
     "date_read": "2026-08-01"
    },
    {
     "id": "17",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-17",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agility's revenue chart scales an installed base from 2,500 to 15,000 Digit v5 and marks itself illustrative",
     "quote": "\"$0 $250 $500 $750 $1,000 $1,250 $1,500 2,500 5,000 7,500 10,000 12,500 15,000 Illustrative annual RaaS revenue(1) … Installed base (total # of Digit v5 deployed)(2)\" with note \"(2) For illustrative purposes only.\"",
     "removal_reason": "",
     "source": "SEC Form 425, Exhibit 99.2 (Investor Presentation, June 2026)",
     "source_url": "https://www.sec.gov/Archives/edgar/data/2074973/000121390026071290/ea029548401ex99-2.htm",
     "locator": "unit-economics slide, installed-base axis",
     "date_read": "2026-08-01"
    },
    {
     "id": "18",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-18",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "BMW announces a next pilot at Leipzig in high-voltage battery assembly, again describing a humanoid robot in the singular and giving no unit count",
     "quote": "\"The deployment in Leipzig is focusing on testing a multifunctional application of the robot\" in the \"assembly of high-voltage batteries and in component manufacturing\"",
     "removal_reason": "",
     "source": "BMW Group press release, 27 Feb 2026",
     "source_url": "https://www.press.bmwgroup.com/global/article/detail/T0455864EN/bmw-group-to-deploy-humanoid-robots-in-production-in-germany-for-the-first-time?language=en",
     "locator": "Leipzig section",
     "date_read": "2026-08-01"
    },
    {
     "id": "19",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-19",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "AGIBOT announced its 10,000th humanoid robot as a production milestone",
     "quote": "\"announced the rollout of its 10,000th humanoid robot\"",
     "removal_reason": "",
     "source": "AGIBOT press release via PR Newswire, 30 Mar 2026",
     "source_url": "https://www.prnewswire.com/news-releases/agibot-reaches-10-000-units-as-real-world-demand-for-robots-accelerates-302728295.html",
     "locator": "opening",
     "date_read": "2026-08-01"
    },
    {
     "id": "20",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-20",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "AGIBOT gives no number for how many of the 10,000 are in real-world use, only an adjective",
     "quote": "\"Of the 10,000 humanoid robots produced, a significant portion is already active in real-world environments\"",
     "removal_reason": "",
     "source": "AGIBOT press release via PR Newswire, 30 Mar 2026",
     "source_url": "https://www.prnewswire.com/news-releases/agibot-reaches-10-000-units-as-real-world-demand-for-robots-accelerates-302728295.html",
     "locator": "deployment paragraph",
     "date_read": "2026-08-01"
    },
    {
     "id": "21",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-21",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "IFR counts 542,000 industrial robots installed in 2024",
     "quote": "\"542,000 robots installed in 2024\"",
     "removal_reason": "",
     "source": "International Federation of Robotics, World Robotics 2025 press release",
     "source_url": "https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years",
     "locator": "installations figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "22",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-22",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "IFR separately counts 4,664,000 industrial robots in operational stock, a different number from installations",
     "quote": "\"The total number of industrial robots in operational use worldwide was 4,664,000 units in 2024\"",
     "removal_reason": "",
     "source": "International Federation of Robotics, World Robotics 2025 press release",
     "source_url": "https://ifr.org/ifr-press-releases/news/global-robot-demand-in-factories-doubles-over-10-years",
     "locator": "operational stock figure",
     "date_read": "2026-08-01"
    },
    {
     "id": "23",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-23",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Unitree's prospectus shows about 9 percent of humanoid robot revenue came from real industrial applications, with about 74 percent from research and education",
     "quote": "\"scientific research and education accounted for 73.6% of humanoid robot revenue, commercial consumption accounted for 17.39%, and real industrial applications only accounted for 9%\"",
     "removal_reason": "",
     "source": "Unitree IPO prospectus, reported by TechFlow and corroborated by a second outlet; prospectus itself not opened",
     "source_url": "https://www.techflowpost.com/en-US/article/31730",
     "locator": "prospectus revenue split, first three quarters of 2025",
     "date_read": "2026-08-01"
    },
    {
     "id": "24",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-24",
     "status": "REPORTED",
     "status_label": "Reported",
     "removed": false,
     "claim": "Tesla has disclosed no Optimus production unit count, and says initial units are for internal data collection rather than customers",
     "quote": "\"training data collection and further functionality development\"",
     "removal_reason": "",
     "source": "Tesla Q2 2026 earnings materials, as reported; Tesla's own deck not opened directly",
     "source_url": "https://www.techtimes.com/articles/321012/20260720/tesla-optimus-production-count-remains-zero-q2-earnings-call-looms-wednesday.htm",
     "locator": "Optimus outlook",
     "date_read": "2026-08-01"
    },
    {
     "id": "25",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-25",
     "status": "EXCLUDED",
     "status_label": "Excluded",
     "removed": true,
     "claim": "Two robots at BMW Spartanburg",
     "quote": "",
     "removal_reason": "Secondary sources only. BMW's own release uses the singular and gives no count. Writing it would commit the error the piece prosecutes.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "26",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-26",
     "status": "EXCLUDED",
     "status_label": "Excluded",
     "removed": true,
     "claim": "Figure has ~740 robots deployed",
     "quote": "",
     "removal_reason": "The primary (Adcock, X, 20 June 2026) returned HTTP 402 and could not be opened. The wording that circulates describes Figure's own premises, not customer deployments. Excluded from the prose entirely — do not reintroduce it as \"high hundreds\" either.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "27",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-27",
     "status": "EXCLUDED",
     "status_label": "Excluded",
     "removed": true,
     "claim": "Agility discloses zero robot counts",
     "quote": "",
     "removal_reason": "FALSE, corrected 2026-08-01 after a review. The press-release exhibit has none, but the investor-presentation exhibit of the same filing discloses 1,000 (as orders). Original draft asserted zero because only Exhibit 99.1 was read.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "28",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-28",
     "status": "EXCLUDED",
     "status_label": "Excluded",
     "removed": true,
     "claim": "About 3 to 4 percent of humanoid shipments do operational work",
     "quote": "",
     "removal_reason": "TechFlow's own derivation from the 9 percent industrial share, not a prospectus disclosure. Excluded as a derived number presented as a disclosed one.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    },
    {
     "id": "29",
     "url": "https://durabilitycurve.com/claims/your-robot-coworker-is-still-a-pilot/#c-29",
     "status": "EXCLUDED",
     "status_label": "Excluded",
     "removed": true,
     "claim": "5,500 shipped implies ~495 industrial units",
     "quote": "",
     "removal_reason": "Cross-basis: 9 percent is a revenue share over three quarters, 5,500 is an annual unit count. Multiplying them is the error the piece is about.",
     "source": "",
     "source_url": "",
     "locator": "",
     "date_read": ""
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Difficulty You're Escaping Was Making You",
    "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/",
    "essay": {
     "title": "The Difficulty You're Escaping Was Making You",
     "url": "https://durabilitycurve.com/blog/difficulty-was-making-you/",
     "published": "2026-07-17"
    },
    "last_checked": "2026-10-04",
    "format": "current",
    "counts": {
     "claims": 9,
     "by_status": {
      "VERIFIED": 9
     },
     "verified_at_primary": 9,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/difficulty-was-making-you/"
   },
   "rows": [
    {
     "id": "DW1",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "To qualify as a licensed London taxi driver, a trainee must learn the layout of roughly 25,000 streets within a 6-mile radius of Charing Cross (the Knowledge), and acquiring it typically takes between 3 and 4 years.",
     "quote": "learn the complex and irregular layout of London's ∼25,000 streets … within a 6-mile radius of Charing Cross … typically takes between 3 and 4 years",
     "removal_reason": "",
     "source": "Katherine Woollett and Eleanor A. Maguire, Acquiring 'the Knowledge' of London's Layout Drives Structural Brain Changes, Current Biology 21(24), 2011, pp. 2109–2114",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC3268356/",
     "locator": "Results and Discussion, opening paragraph · Highlights, first bullet (3–4 years) · Figure 1 caption (6-mile radius)",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW2",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Scanned at the start of training and again 3–4 years later, the trainee taxi drivers who qualified showed an increase in posterior hippocampal grey matter, while trainees who failed to qualify and non-driver controls showed no structural brain change. Because the same people were scanned before and after, this longitudinal result is what supports reading the change as produced by the training.",
     "quote": "Those who qualified experienced increased gray matter in posterior hippocampus … No structural brain changes were observed in trainees who failed to qualify or control participants",
     "removal_reason": "",
     "source": "Katherine Woollett and Eleanor A. Maguire, Acquiring 'the Knowledge' of London's Layout Drives Structural Brain Changes, Current Biology 21(24), 2011, pp. 2109–2114",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC3268356/",
     "locator": "Highlights, third bullet · Summary, penultimate sentence · Figure 2 caption",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW3",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Compared with London bus drivers, who were matched for driving experience and stress but follow constrained routes, London taxi drivers had less grey matter volume in the anterior hippocampi and were worse at acquiring new visuo-spatial information. Scope: the comparison is cross-sectional, and the functional difference rests on one test, delayed recall of the Rey-Osterrieth Complex Figure (a line drawing); no new-environment or route-learning test was run.",
     "quote": "taxi drivers had … less volume in anterior hippocampi … the ability to acquire new visuo-spatial information was worse in taxi drivers than in bus drivers",
     "removal_reason": "",
     "source": "Eleanor A. Maguire, Katherine Woollett and Hugo J. Spiers, London Taxi Drivers and Bus Drivers: A Structural MRI and Neuropsychological Analysis, Hippocampus 16(12), 2006, pp. 1091–1101",
     "source_url": "https://doi.org/10.1002/hipo.20233",
     "locator": "Abstract · Results, neuropsychological comparison and Table 2, p. 1095 of the authors' copy (Rey-Osterrieth Complex Figure, delayed recall)",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW4",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Among regular drivers, those with greater lifetime GPS experience had worse spatial memory when they had to navigate without GPS; in a small follow-up of a subset of participants three years later, greater GPS use since the first test was associated with a steeper decline in hippocampal-dependent spatial memory. Scope: both findings are correlational, and the authors call the longitudinal sample small.",
     "quote": "greater lifetime GPS experience have worse spatial memory during self-guided navigation … greater GPS use since initial testing was associated with a steeper decline",
     "removal_reason": "",
     "source": "Louisa Dahmani and Véronique D. Bohbot, Habitual use of GPS negatively impacts spatial memory during self-guided navigation, Scientific Reports 10, article 6310, 2020",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC7156656/",
     "locator": "Abstract, fourth and sixth sentences (the follow-up retested 13 of the 50 participants)",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW5",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Students who learned prose passages by taking free-recall tests retained substantially more on delayed final tests than students who restudied the passages, even though repeated studying made students more confident that they would remember the material.",
     "quote": "on the delayed tests, prior testing produced substantially greater retention than studying, even though repeated studying increased students' confidence",
     "removal_reason": "",
     "source": "Henry L. Roediger III and Jeffrey D. Karpicke, Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention, Psychological Science 17(3), 2006, pp. 249–255",
     "source_url": "https://doi.org/10.1111/j.1467-9280.2006.01693.x",
     "locator": "Abstract, p. 249 · confidence measure: Experiment 2, Questionnaire, p. 252",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW6",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Barcaui's randomised controlled trial with undergraduates learning about AI, students who used ChatGPT as a study aid scored 57.5% on a surprise retention test 45 days later, against 68.5% for students who used traditional, non-AI study methods.",
     "quote": "surprise test 45 days after learning … ChatGPT scored significantly lower on the retention test (57.5 % correct) compared to those who studied traditionally (68.5 % correct)",
     "removal_reason": "",
     "source": "André Barcaui, ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention, Social Sciences & Humanities Open 12, article 102287, 2025 (preprint SSRN 5353041, posted 21 July 2025)",
     "source_url": "https://doi.org/10.1016/j.ssaho.2025.102287",
     "locator": "Abstract, fourth and fifth sentences · the SSRN preprint abstract carries the same result sentence",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW7",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a preliminary essay-writing study posted to arXiv, participants who wrote with an LLM showed the weakest EEG brain connectivity of three groups (LLM, search engine, no tools), reported the lowest sense of ownership of their essays, and struggled to quote their own work accurately. Scope: an arXiv preprint, not a peer-reviewed paper.",
     "quote": "LLM users displayed the weakest connectivity … Self-reported ownership of essays was the lowest in the LLM group … LLM users also struggled to accurately quote their own work",
     "removal_reason": "",
     "source": "Nataliya Kosmyna et al., Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, arXiv:2506.08872",
     "source_url": "https://arxiv.org/abs/2506.08872",
     "locator": "Abstract (v2, revised 31 December 2025)",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW8",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In a cross-sectional MRI study, the posterior hippocampi of licensed London taxi drivers were significantly larger than those of control subjects who did not drive taxis, and hippocampal volume correlated with the time spent as a taxi driver, positively in the posterior hippocampus.",
     "quote": "posterior hippocampi of taxi drivers were significantly larger … Hippocampal volume correlated with the amount of time spent as a taxi driver (positively in the posterior …)",
     "removal_reason": "",
     "source": "Eleanor A. Maguire et al., Navigation-related structural change in the hippocampi of taxi drivers, Proceedings of the National Academy of Sciences 97(8), 2000, pp. 4398–4403",
     "source_url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC18253/",
     "locator": "Abstract",
     "date_read": "2026-10-04"
    },
    {
     "id": "DW9",
     "url": "https://durabilitycurve.com/claims/difficulty-was-making-you/#c-DW9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Experiment 2, students who repeatedly tested themselves on a passage recalled 61% of it a week later, against 40% for students who repeatedly restudied it, although the self-testers had read the passage 3.4 times and the restudiers 14.2 times.",
     "quote": "repeated-testing condition recalled much more after a week … (61% vs. 40%) … only 3.4 times and those in the latter condition read it 14.2 times",
     "removal_reason": "",
     "source": "Henry L. Roediger III and Jeffrey D. Karpicke, Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention, Psychological Science 17(3), 2006, pp. 249–255 (authors' copy)",
     "source_url": "https://learninglab.psych.purdue.edu/downloads/2006/2006_Roediger_Karpicke_PsychSci.pdf",
     "locator": "General Discussion, p. 253 · Experiment 2, Final Tests, p. 252 · Table 1 (reading counts)",
     "date_read": "2026-10-04"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Your Research Agent Cites Sources It Never Read",
    "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/",
    "essay": {
     "title": "Your Research Agent Cites Sources It Never Read",
     "url": "https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/",
     "published": "2026-07-11"
    },
    "last_checked": "2026-10-06",
    "format": "current",
    "counts": {
     "claims": 11,
     "by_status": {
      "VERIFIED": 11
     },
     "verified_at_primary": 11,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/"
   },
   "rows": [
    {
     "id": "RA1",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Proof-of-Use authors define tool-call hacking: an RL-trained research agent earns high reward by satisfying tool-call format and answer-correctness objectives while its reasoning does not causally depend on what the tools return.",
     "quote": "satisfying tool-call format requirements … its stepwise reasoning and action behavior does not causally depend on the content returned by the tools",
     "removal_reason": "",
     "source": "Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026)",
     "source_url": "https://arxiv.org/abs/2510.10931",
     "locator": "§3.1, \"Tool-Call Hacking (Definition)\"",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA2",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Under tool-call hacking, an agent's tool use collapses to a narrow or repetitive pattern, overusing a single tool or a fixed call template.",
     "quote": "collapses to a narrow or repetitive calling pattern … the policy tends to overuse a single tool or a fixed call template",
     "removal_reason": "",
     "source": "Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026)",
     "source_url": "https://arxiv.org/abs/2510.10931",
     "locator": "§3.1, \"Manifestations of Tool-Call Hacking\"",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA3",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Because the link between retrieved evidence and reasoning is weakly observable, agents can maximise surface-level reward signals without grounding their reasoning in the returned evidence, which shows up as mode collapse through tool overuse.",
     "quote": "maximize surface-level reward signals without genuinely grounding their reasoning in the returned evidence … mode collapse via tool overuse",
     "removal_reason": "",
     "source": "Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026)",
     "source_url": "https://arxiv.org/abs/2510.10931",
     "locator": "Abstract",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA4",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Proof-of-Use's fix is a multi-part reward: two step-level rewards that constrain citation validity, plus an answer-support alignment reward for consistency between the final answer and the retrieved evidence. The final reward averages the three.",
     "quote": "two progressive process rewards that constrain citation validity at intermediate steps … a global Answer–Support Alignment reward that enforces consistency between final answers and retrieved evidence",
     "removal_reason": "",
     "source": "Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026)",
     "source_url": "https://arxiv.org/abs/2510.10931",
     "locator": "Abstract · §3.3.4 Overall Objective, Eq. 10 (the average of the citation, evidence-sensitive and answer terms)",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA5",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "One of Proof-of-Use's step rewards is perturbation-based: the cited evidence is replaced with unrelated text, and a model that truly relies on that evidence should become less confident it was helpful.",
     "quote": "replacing their contents with topic-unrelated snippets … A model that truly relies on the cited evidence should decrease its helpfulness confidence",
     "removal_reason": "",
     "source": "Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026)",
     "source_url": "https://arxiv.org/abs/2510.10931",
     "locator": "§3.3.2 Evidence-Sensitive Reward (same mechanism in v1, 13 October 2025)",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA6",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Needham et al. built a benchmark of 1,000 prompts and transcripts from 61 datasets. Gemini-2.5-Pro told evaluation from real-world deployment with an AUC of 0.83, below a simple human baseline of 0.92.",
     "quote": "1,000 prompts and transcripts from 61 distinct datasets … Gemini-2.5-Pro reaches an AUC of 0.83 … do not yet surpass our simple human baseline (AUC of 0.92)",
     "removal_reason": "",
     "source": "Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025)",
     "source_url": "https://arxiv.org/abs/2505.23836",
     "locator": "Abstract",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA7",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Both the models and the humans in Needham et al. were better at spotting evaluations in agentic settings than in chat settings.",
     "quote": "both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings",
     "removal_reason": "",
     "source": "Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025)",
     "source_url": "https://arxiv.org/abs/2505.23836",
     "locator": "Abstract",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA8",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Goodfire's study, Gemini 3.1 Pro identified the evaluation it was being tested on, described its format and cited a paper critiquing it; the benchmark was StereoSet.",
     "quote": "Gemini 3.1 Pro not only identified the eval it was being tested on, but also described its format and cited a paper critiquing it",
     "removal_reason": "",
     "source": "Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026",
     "source_url": "https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety",
     "locator": "Introduction · §2 (the StereoSet critique by Blodgett et al., Appendix B)",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA9",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Of the 515 manually verified instances of verbalised eval awareness Goodfire found, across eight models and 19 benchmarks, 95% appeared only in the chain of thought, 4% only in the response and 1% in both.",
     "quote": "Across the 515 instances, 95% appear exclusively in the chain of thought, 4% in the response only, and 1% in both.",
     "removal_reason": "",
     "source": "Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026",
     "source_url": "https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety",
     "locator": "§1, \"Verbalized eval awareness appears much more often in the chain of thought than in the response\" · Summary of results (19 benchmarks, 8 models, 515 instances)",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA10",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On the Fortress benchmark, four models other than the Gemini models refused harmful requests 3 to 18 percentage points more often in runs where they verbalised eval awareness.",
     "quote": "with four other models on the Fortress benchmark … models refuse harmful requests 3–18 percentage points more often when they verbalize eval awareness",
     "removal_reason": "",
     "source": "Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026",
     "source_url": "https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety",
     "locator": "Summary of results, second item · §2, opening paragraph",
     "date_read": "2026-10-06"
    },
    {
     "id": "RA11",
     "url": "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Knecht, Florin and Hagendorff found that spontaneously occurring verbalised eval awareness shifted open-weight reasoning models' answer distributions by at most 3.7 percentage points, and conclude it may pose a smaller safety risk than the literature assumes.",
     "quote": "spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points … Evaluation awareness may pose a smaller safety risk than the current literature assumes",
     "removal_reason": "",
     "source": "Amelie Knecht, Lucas Florin and Thilo Hagendorff, Evaluation Awareness in Language Models Has Limited Effect on Behaviour, arXiv:2605.05835 (2026)",
     "source_url": "https://arxiv.org/abs/2605.05835",
     "locator": "Abstract",
     "date_read": "2026-10-06"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Your Multi-Agent System Is an Org Chart",
    "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/",
    "essay": {
     "title": "Your Multi-Agent System Is an Org Chart",
     "url": "https://durabilitycurve.com/blog/your-multi-agent-system-is-an-org-chart/",
     "published": "2026-07-07"
    },
    "last_checked": "2026-10-03",
    "format": "current",
    "counts": {
     "claims": 9,
     "by_status": {
      "VERIFIED": 9
     },
     "verified_at_primary": 9,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/"
   },
   "rows": [
    {
     "id": "MA1",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cognition's June 2025 post, written from its own experience building coding agents and explicitly dated to 2025, said running multiple agents in collaboration produced fragile systems because decision-making ended up too dispersed and context could not be shared thoroughly enough between the agents.",
     "quote": "in 2025, running multiple agents in collaboration only results in fragile systems … decision-making … too dispersed and context isn’t able to be shared thoroughly enough",
     "removal_reason": "",
     "source": "Walden Yan, Cognition, Don't Build Multi-Agents, 12 June 2025",
     "source_url": "https://cognition.com/blog/dont-build-multi-agents",
     "locator": "section \"Applying the Principles\", sub-section \"Multi-Agents\", final paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA2",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On Anthropic's own internal research eval, not an external benchmark, its multi-agent research system (a Claude Opus 4 lead agent with Claude Sonnet 4 subagents) outperformed single-agent Claude Opus 4 by 90.2%.",
     "quote": "Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval",
     "removal_reason": "",
     "source": "Anthropic Engineering, How we built our multi-agent research system, 13 June 2025",
     "source_url": "https://www.anthropic.com/engineering/multi-agent-research-system",
     "locator": "section \"Benefits of a multi-agent system\", second paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA3",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On the BrowseComp browsing evaluation specifically, Anthropic found that token usage by itself explained 80% of the performance variance; three factors together explained 95%.",
     "quote": "three factors explained 95% of the performance variance in the BrowseComp evaluation … token usage by itself explains 80% of the variance",
     "removal_reason": "",
     "source": "Anthropic Engineering, How we built our multi-agent research system, 13 June 2025",
     "source_url": "https://www.anthropic.com/engineering/multi-agent-research-system",
     "locator": "section \"Benefits of a multi-agent system\", third paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA4",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Anthropic's own data, multi-agent systems used about 15× more tokens than chats, against about 4× for single agents.",
     "quote": "In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats",
     "removal_reason": "",
     "source": "Anthropic Engineering, How we built our multi-agent research system, 13 June 2025",
     "source_url": "https://www.anthropic.com/engineering/multi-agent-research-system",
     "locator": "section \"Benefits of a multi-agent system\", fourth paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA5",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic wrote in June 2025 that domains requiring all agents to share the same context, or involving many dependencies between agents, are not a good fit for multi-agent systems today, and that most coding tasks involve fewer truly parallelisable tasks than research.",
     "quote": "share the same context or involve many dependencies between agents are not a good fit … most coding tasks involve fewer truly parallelizable tasks than research",
     "removal_reason": "",
     "source": "Anthropic Engineering, How we built our multi-agent research system, 13 June 2025",
     "source_url": "https://www.anthropic.com/engineering/multi-agent-research-system",
     "locator": "section \"Benefits of a multi-agent system\", fourth paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA6",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "A preprint by Tran and Kiela found that, with reasoning-token budgets held equal, single-agent systems consistently match or outperform multi-agent systems on multi-hop reasoning tasks across three model families (Qwen3, DeepSeek-R1-Distill-Llama and Gemini 2.5). The finding covers multi-hop reasoning only.",
     "quote": "three model families (Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5) … SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant",
     "removal_reason": "",
     "source": "Dat Tran and Douwe Kiela, Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets, arXiv preprint 2604.02460v2",
     "source_url": "https://arxiv.org/abs/2604.02460",
     "locator": "Abstract (revised 11 April 2026)",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA7",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude Code's current sub-agents documentation lists its built-in Explore and Plan subagents with read-only tools, Write and Edit denied.",
     "quote": "Explore … Tools: read-only tools; Write and Edit are denied … Plan … Tools: read-only tools; Write and Edit are denied",
     "removal_reason": "",
     "source": "Claude Code Docs, Create custom subagents",
     "source_url": "https://code.claude.com/docs/en/sub-agents",
     "locator": "section \"Built-in subagents\", Explore and Plan tabs, Tools line",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA8",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cognition's April 2026 follow-up said its original observations still hold for parallel-writer swarms, and that the multi-agent setups that work in its practice are those where multiple agents contribute intelligence while writes stay single-threaded.",
     "quote": "Our original observations still hold today for parallel-writer swarms … setups where multiple agents contribute intelligence to a task while writes stay single-threaded",
     "removal_reason": "",
     "source": "Walden Yan, Cognition, Multi-Agents: What's Actually Working, 22 April 2026",
     "source_url": "https://cognition.com/blog/multi-agents-working",
     "locator": "introduction, second paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "MA9",
     "url": "https://durabilitycurve.com/claims/your-multi-agent-system-is-an-org-chart/#c-MA9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Cognition states, as its own company figure, that Devin usage in its largest-enterprises segment, which it describes as traditionally cautious toward new technologies, grew roughly 8× over the six months before its April 2026 post.",
     "quote": "Devin usage in our largest enterprises segment … traditionally been cautious toward adopting new technologies, we see an explosion over the last 6 months (~8x)",
     "removal_reason": "",
     "source": "Walden Yan, Cognition, Multi-Agents: What's Actually Working, 22 April 2026",
     "source_url": "https://cognition.com/blog/multi-agents-working",
     "locator": "section \"What Changed in the Last 10 Months\", first paragraph",
     "date_read": "2026-10-03"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: The Seven-Layer Agent Audit",
    "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/",
    "essay": {
     "title": "The Seven-Layer Agent Audit",
     "url": "https://durabilitycurve.com/blog/the-seven-layer-agent-audit/",
     "published": "2026-06-17"
    },
    "last_checked": "2026-10-11",
    "format": "current",
    "counts": {
     "claims": 8,
     "by_status": {
      "VERIFIED": 8
     },
     "verified_at_primary": 8,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/the-seven-layer-agent-audit/"
   },
   "rows": [
    {
     "id": "SL1",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "SWE-agent was built by a team at Princeton Language and Intelligence, Princeton University, and was published in the NeurIPS 2024 Main Conference Track.",
     "quote": "Princeton Language and Intelligence, Princeton University … Advances in Neural Information Processing Systems 37 (NeurIPS 2024) Main Conference Track",
     "removal_reason": "",
     "source": "John Yang, Carlos E. Jimenez, Alexander Wettig, Kilian Lieret, Shunyu Yao, Karthik Narasimhan and Ofir Press, SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, NeurIPS 2024 (arXiv:2405.15793)",
     "source_url": "https://arxiv.org/abs/2405.15793",
     "locator": "Author affiliations, title block (arXiv HTML v3) · NeurIPS 2024 proceedings page, citation line",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL2",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Using GPT-4 Turbo, SWE-agent resolved 12.47% of the 2,294 tasks in the SWE-bench test set, against a previous best of 3.8% by a non-interactive, retrieval-augmented system.",
     "quote": "SWE-agent solves 12.47% of the 2,294 SWE-bench test tasks … previous best resolve rate of 3.8% by a non-interactive, retrieval-augmented system",
     "removal_reason": "",
     "source": "John Yang, Carlos E. Jimenez et al., SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, NeurIPS 2024 (arXiv:2405.15793)",
     "source_url": "https://arxiv.org/abs/2405.15793",
     "locator": "§1 Introduction · §5 Results (286 of 2,294)",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL3",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The 3.8% previous best was the retrieval-augmented baseline running on Claude 3 Opus (3.79% of the full test set); the same baseline on GPT-4 Turbo resolved 1.31%.",
     "quote": "RAG w/ GPT-4 Turbo 1.31 … w/ Claude 3 Opus 3.79",
     "removal_reason": "",
     "source": "John Yang, Carlos E. Jimenez et al., SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, NeurIPS 2024 (arXiv:2405.15793)",
     "source_url": "https://arxiv.org/abs/2405.15793",
     "locator": "Table 1, SWE-bench column, % Resolved (same values in v1, v2 and v3)",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL4",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "With the model held at GPT-4 Turbo on the 300-task SWE-bench Lite split, SWE-agent resolved 18.00% of tasks against 11.00% for an agent working through the plain Linux shell, a 64% relative increase.",
     "quote": "An LM-friendly ACI's value is confirmed by SWE-agent's 64% relative increase compared to Shell-only, both with GPT-4 Turbo.",
     "removal_reason": "",
     "source": "John Yang, Carlos E. Jimenez et al., SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, NeurIPS 2024 (arXiv:2405.15793)",
     "source_url": "https://arxiv.org/abs/2405.15793",
     "locator": "§5 Results, first paragraph · Table 1, SWE-bench Lite column (Shell-only agent w/ GPT-4 Turbo 11.00; SWE-agent w/ GPT-4 Turbo 18.00)",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL5",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "SWE-agent's interface changed what the agent could see and do: commands for searching, viewing and editing files, guardrails against common mistakes, and concise feedback on each command's effects.",
     "quote": "a small set of simple actions for viewing, searching through and editing files … guardrails to prevent common mistakes … concise feedback about a command's effects",
     "removal_reason": "",
     "source": "John Yang, Carlos E. Jimenez et al., SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering, NeurIPS 2024 (arXiv:2405.15793)",
     "source_url": "https://arxiv.org/abs/2405.15793",
     "locator": "§1 Introduction · §3, opening paragraph (search/navigation, file viewer, file editor, context management)",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL6",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In SkillsBench, curated skills raised the average pass rate from 33.9% to 50.5%, a gain of 16.6 percentage points across 87 tasks and 18 model-harness configurations; versions 1 to 3 of the paper reported 16.2 points on 84 evaluated tasks (of 86).",
     "quote": "Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain)",
     "removal_reason": "",
     "source": "Xiangyi Li et al., SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, arXiv:2602.12670",
     "source_url": "https://arxiv.org/abs/2602.12670",
     "locator": "Abstract (v4, 14 June 2026) · v1–v3 abstract: \"Curated Skills raise average pass rate by 16.2 percentage points\" · v1 §3.5 Metrics, \"Pass Rate\": \"a fixed denominator of 84 (the number of evaluated tasks)\"",
     "date_read": "2026-10-11"
    },
    {
     "id": "SL7",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Skills that a model wrote for itself gave no average benefit. In versions 1 to 3 they averaged 1.3 points below the no-skills baseline; in version 4 they landed below it on all three configurations tested, by 8.1 to 11.5 points.",
     "quote": "Self-generated Skills land below the no-Skills baseline on all three configurations",
     "removal_reason": "",
     "source": "Xiangyi Li et al., SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, arXiv:2602.12670",
     "source_url": "https://arxiv.org/abs/2602.12670",
     "locator": "v4 §5.1.1, paragraph \"Self-generated Skills do not substitute for curated ones\" (−8.1, −11.3 and −11.5 pp) · v1 §4.1.1, Finding 3 (–1.3pp on average)",
     "date_read": "2026-10-06"
    },
    {
     "id": "SL8",
     "url": "https://durabilitycurve.com/claims/the-seven-layer-agent-audit/#c-SL8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Language models use long contexts unevenly: performance is often highest when the relevant information is at the beginning or end of the input, and degrades significantly when it sits in the middle.",
     "quote": "often highest when relevant information occurs at the beginning or end … significantly degrades when models must access relevant information in the middle of long contexts",
     "removal_reason": "",
     "source": "Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni and Percy Liang, Lost in the Middle: How Language Models Use Long Contexts, Transactions of the Association for Computational Linguistics 12, 2024, pp. 157–173",
     "source_url": "https://aclanthology.org/2024.tacl-1.9/",
     "locator": "Abstract",
     "date_read": "2026-10-06"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Skills Are Package Management for Your AI",
    "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/",
    "essay": {
     "title": "Skills Are Package Management for Your AI",
     "url": "https://durabilitycurve.com/blog/skills-are-package-management-for-your-ai/",
     "published": "2026-06-07"
    },
    "last_checked": "2026-10-02",
    "format": "current",
    "counts": {
     "claims": 8,
     "by_status": {
      "VERIFIED": 8
     },
     "verified_at_primary": 8,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/"
   },
   "rows": [
    {
     "id": "S1",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "SkillsMP's live index contained 3,337,075 collected public SKILL.md files on 2 October 2026",
     "quote": "3,337,075 Collected SKILL.md files",
     "removal_reason": "",
     "source": "SkillsMP, Agent Skills Marketplace",
     "source_url": "https://skillsmp.com/",
     "locator": "homepage hero count",
     "date_read": "2026-10-02"
    },
    {
     "id": "S2",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic published Agent Skills as an open standard intended for cross-platform portability",
     "quote": "We've published Agent Skills as an open standard for cross-platform portability.",
     "removal_reason": "",
     "source": "Anthropic, Equipping agents for the real world with Agent Skills",
     "source_url": "https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills",
     "locator": "18 December 2025 update",
     "date_read": "2026-10-02"
    },
    {
     "id": "S3",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The Agent Skills format uses SKILL.md instructions that an agent loads when a task matches the skill description",
     "quote": "When a task matches a skill’s description, the agent reads the full SKILL.md instructions into context.",
     "removal_reason": "",
     "source": "Agent Skills, Overview",
     "source_url": "https://agentskills.io/home",
     "locator": "How do Agent Skills work? — Activation",
     "date_read": "2026-10-02"
    },
    {
     "id": "S4",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Agent Skills are supported across a broad set of AI tools and agentic clients",
     "quote": "Agent Skills are supported by a large number of AI tools and agentic clients",
     "removal_reason": "",
     "source": "Agent Skills, Overview and official Client Showcase registry",
     "source_url": "https://agentskills.io/home",
     "locator": "Where can I use Agent Skills?; linked client registry counted 46 entries on 2026-10-02",
     "date_read": "2026-10-02"
    },
    {
     "id": "S5",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Recent security research frames third-party agent skills as a software-supply-chain risk because developers may install and execute them with implicit trust",
     "quote": "developers install and execute third-party skills with implicit trust and no formal analysis",
     "removal_reason": "",
     "source": "Bhardwaj, Formal Analysis and Supply Chain Security for Agentic AI Skills, arXiv:2603.00195v2",
     "source_url": "https://arxiv.org/html/2603.00195v2",
     "locator": "§1 Introduction, lines 205–206",
     "date_read": "2026-10-02"
    },
    {
     "id": "S6",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The same paper defines an Agent Skill Bill of Materials using CycloneDX and records identity/version/hash, capabilities, dependencies, trust metadata and vulnerability status",
     "quote": "An ASBOM is a CycloneDX 1.6-compatible document that records",
     "removal_reason": "",
     "source": "Bhardwaj, Formal Analysis and Supply Chain Security for Agentic AI Skills, arXiv:2603.00195v2",
     "source_url": "https://arxiv.org/html/2603.00195v2",
     "locator": "§6.6, Definition 6.8, lines 965–989",
     "date_read": "2026-10-02"
    },
    {
     "id": "S7",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "SkillsBench's current release reports that curated Skills raise mean resolution rate from 33.9% to 50.5%, a 16.6-point lift, across the 18-configuration paper aggregate",
     "quote": "In the 18-configuration paper aggregate … The mean resolution rate rises from 33.9% to 50.5% (+16.6 points",
     "removal_reason": "",
     "source": "SkillsBench 1.1: Agent Skills Benchmark Release",
     "source_url": "https://www.skillsbench.ai/blogs/skillsbench-1-1",
     "locator": "Results",
     "date_read": "2026-10-02"
    },
    {
     "id": "S8",
     "url": "https://durabilitycurve.com/claims/skills-are-package-management-for-your-ai/#c-S8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "On SkillsBench's three dedicated self-generated-Skills configurations, self-generated Skills underperformed the no-Skills baseline by 8.1 to 11.5 points",
     "quote": "Self-generated Skills change resolution rate by −8.1 points for Claude Code + Opus 4.7, −11.3 points for Codex + GPT-5.5, and −11.5 points",
     "removal_reason": "",
     "source": "SkillsBench 1.1: Agent Skills Benchmark Release",
     "source_url": "https://www.skillsbench.ai/blogs/skillsbench-1-1",
     "locator": "Self-Generated Vs. Curated Skills",
     "date_read": "2026-10-02"
    }
   ]
  },
  {
   "ledger": {
    "title": "Claim Ledger: Same Model, Different Product: The Case for Harness Engineering",
    "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/",
    "essay": {
     "title": "Same Model, Different Product: The Case for Harness Engineering",
     "url": "https://durabilitycurve.com/blog/harness-engineering-same-model-different-product/",
     "published": "2026-05-30"
    },
    "last_checked": "2026-10-03",
    "format": "current",
    "counts": {
     "claims": 14,
     "by_status": {
      "VERIFIED": 14
     },
     "verified_at_primary": 14,
     "removed": 0
    },
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "The Durability Curve (Harry Floyd), https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/"
   },
   "rows": [
    {
     "id": "HE1",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE1",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the Meta-Harness paper's TerminalBench-2 table, Claude Opus 4.6 has pass rates of 58.0 under Claude Code (an official-leaderboard result the paper reports) and 76.4 under the Meta-Harness harness. Scope: Meta-Harness searched and was finally evaluated on the same 89-task TerminalBench-2 benchmark, and the paper describes the resulting harness as specialised to the TerminalBench-2 regime.",
     "quote": "search and final evaluation on the same 89-task benchmark … specialized to the TerminalBench-2 regime … Claude Opus 4.6 … Claude Code × 58.0 … Meta-Harness ✓ 76.4",
     "removal_reason": "",
     "source": "Lee, Nair, Zhang, Lee, Khattab and Finn, Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv:2603.28052v1",
     "source_url": "https://arxiv.org/pdf/2603.28052",
     "locator": "p. 9, §4.3 Evaluating Agentic Coding Harnesses on TerminalBench-2 (setup paragraph) · Table 7, Claude Opus 4.6 block; caption: results of others are from the official leaderboard",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE2",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE2",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the same table, Claude Haiku 4.5 has pass rates of 27.5 under Claude Code (an official-leaderboard result the paper reports) and 37.6 under the Meta-Harness harness. Scope: Meta-Harness searched and was finally evaluated on the same 89-task TerminalBench-2 benchmark, and the paper describes the resulting harness as specialised to the TerminalBench-2 regime.",
     "quote": "search and final evaluation on the same 89-task benchmark … specialized to the TerminalBench-2 regime … Claude Haiku 4.5 … Claude Code × 27.5 … Meta-Harness ✓ 37.6",
     "removal_reason": "",
     "source": "Lee, Nair, Zhang, Lee, Khattab and Finn, Meta-Harness: End-to-End Optimization of Model Harnesses, arXiv:2603.28052v1",
     "source_url": "https://arxiv.org/pdf/2603.28052",
     "locator": "p. 9, §4.3 (setup paragraph) · Table 7, Claude Haiku 4.5 block; also Figure 1 (right), p. 1",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE3",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE3",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "LangChain reports improving its coding agent deepagents-cli by 13.7 points on Terminal Bench 2.0, from 52.8 to 66.5, by changing only the harness while keeping the model fixed at GPT-5.2-Codex.",
     "quote": "improve deepagents-cli … 13.7 points from 52.8 to 66.5 on Terminal Bench 2.0. We only tweaked the harness and kept the model fixed, gpt-5.2-codex",
     "removal_reason": "",
     "source": "LangChain (Vivek Trivedy), Improving Deep Agents with harness engineering, 17 February 2026",
     "source_url": "https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering",
     "locator": "opening section, paragraph beginning \"We used a simple recipe\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE4",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE4",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude Code's command reference documents /compact as a command that frees up context by summarising the conversation so far.",
     "quote": "/compact [instructions] … Free up context by summarizing the conversation so far.",
     "removal_reason": "",
     "source": "Claude Code Docs, Commands",
     "source_url": "https://code.claude.com/docs/en/commands",
     "locator": "command table, row \"/compact [instructions]\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE5",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE5",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude Code's environment-variable reference documents CLAUDE_CODE_MAX_RETRIES, which overrides the number of times Claude Code retries failed API requests.",
     "quote": "CLAUDE_CODE_MAX_RETRIES … Override the number of times to retry failed API requests",
     "removal_reason": "",
     "source": "Claude Code Docs, Environment variables",
     "source_url": "https://code.claude.com/docs/en/env-vars",
     "locator": "Variables list, row \"CLAUDE_CODE_MAX_RETRIES\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE6",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE6",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Claude Code's features overview describes subagents as isolated workers with their own context, running separately from the main conversation.",
     "quote": "Subagents … are isolated workers that run separately from your main conversation … Isolated worker with its own context",
     "removal_reason": "",
     "source": "Claude Code Docs, Extend Claude Code (features overview)",
     "source_url": "https://code.claude.com/docs/en/features-overview",
     "locator": "Compare similar features, tab \"Skill vs Subagent\", row \"What it is\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE7",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE7",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Manus warns, from its own experience building an agent rather than from a measured experiment, that when users plug hundreds of tools into an agent's action space, the model is more likely to select the wrong action or take an inefficient path.",
     "quote": "hundreds of mysterious tools … the model is more likely to select the wrong action or take an inefficient path",
     "removal_reason": "",
     "source": "Yichao 'Peak' Ji, Manus, Context Engineering for AI Agents: Lessons from Building Manus, 18 July 2025",
     "source_url": "https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus",
     "locator": "section \"Mask, Don't Remove\", first paragraph",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE8",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE8",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Manus deliberately leaves failed actions and errors in the agent's context, arguing that erasing failure removes the evidence the model needs to adapt.",
     "quote": "Erasing failure removes evidence. And without evidence, the model can't adapt. … leave the wrong turns in the context",
     "removal_reason": "",
     "source": "Yichao 'Peak' Ji, Manus, Context Engineering for AI Agents: Lessons from Building Manus, 18 July 2025",
     "source_url": "https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus",
     "locator": "section \"Keep the Wrong Stuff In\", second and third paragraphs",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE9",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE9",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In Can Bölük's edit-format benchmark, the hashline edit format beat the patch format in 14 of 16 models.",
     "quote": "hashline beats patch in 14/16 models",
     "removal_reason": "",
     "source": "Can Bölük, Stencil, We improved 15 LLMs at coding in one afternoon. Only the harness changed., 12 February 2026",
     "source_url": "https://stencil.so/blog/the-harness-problem",
     "locator": "benchmark chart caption (3 runs × 180 tasks per model)",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE10",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE10",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "In the same benchmark, Grok Code Fast 1 went from 6.7% with the patch edit format to 68.3% with hashline.",
     "quote": "Grok Code Fast 1 went from 6.7% to 68.3%, a tenfold improvement … patch 6.7% … hashline 68.3%",
     "removal_reason": "",
     "source": "Can Bölük, Stencil, We improved 15 LLMs at coding in one afternoon. Only the harness changed., 12 February 2026",
     "source_url": "https://stencil.so/blog/the-harness-problem",
     "locator": "results paragraph beginning \"Sixteen models, three edit tools\" · chart row 01 point labels",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE11",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE11",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Menlo Ventures estimated in December 2025 that Anthropic held 54% of the enterprise coding market, up from 42% six months earlier, driven in large part by Claude Code. The figure comes from the report's LLM market-share analysis, which estimates model-provider share from production API usage.",
     "quote": "coding market … an estimated 54% market share … up from 42% just six months ago, driven in large part by the popularity of Claude Code",
     "removal_reason": "",
     "source": "Menlo Ventures, 2025: The State of Generative AI in the Enterprise, 9 December 2025",
     "source_url": "https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/",
     "locator": "section \"LLM Market Share: Anthropic Extends Its Lead in the Enterprise\", paragraph after the share charts · methodology note \"LLM Market Share\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE12",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE12",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "Anthropic said in February 2026 that Claude Code's run-rate revenue had grown to over $2.5 billion.",
     "quote": "Today, Claude Code’s run-rate revenue has grown to over $2.5 billion",
     "removal_reason": "",
     "source": "Anthropic, Series G funding announcement, 12 February 2026",
     "source_url": "https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation",
     "locator": "paragraph beginning \"Claude Code represents a new era of agentic coding\"",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE13",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE13",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "DeepSeek publishes DeepSeek Harness (dsh) from its deepseek-ai GitHub organisation, and the README describes it as an open-source agent harness developed by DeepSeek AI.",
     "quote": "DeepSeek Harness (dsh) is an open-source agent harness developed by DeepSeek AI.",
     "removal_reason": "",
     "source": "DeepSeek AI, deepseek-ai/deepseek-harness README, GitHub",
     "source_url": "https://github.com/deepseek-ai/deepseek-harness",
     "locator": "README opening paragraph · DeepSeek's own site, deepseek.com/harness, links the repository",
     "date_read": "2026-10-03"
    },
    {
     "id": "HE14",
     "url": "https://durabilitycurve.com/claims/harness-engineering-same-model-different-product/#c-HE14",
     "status": "VERIFIED",
     "status_label": "Verified",
     "removed": false,
     "claim": "The DeepSeek Harness README labels the project a developer preview.",
     "quote": "DeepSeek Harness is in developer preview and iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.",
     "removal_reason": "",
     "source": "DeepSeek AI, deepseek-ai/deepseek-harness README, GitHub",
     "source_url": "https://github.com/deepseek-ai/deepseek-harness",
     "locator": "README section \"Developer preview\"",
     "date_read": "2026-10-03"
    }
   ]
  }
 ]
}