{
 "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"
  }
 ]
}