--- title: "Claim Ledger: Your Research Agent Cites Sources It Never Read" description: "The same trap has killed pricing models and trading desks for decades. One move tells you if your number is next." author: "Harry Floyd" publication: "The Durability Curve" canonical: "https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/" essay: "https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/" substack: "https://harryfloyd.substack.com/p/your-research-agent-cites-sources-it-never-read" published: "2026-07-11" ledger_date: "2026-10-06" law: "Law I" claims: 11 removed_before_publication: 0 --- # Claim Ledger: Your Research Agent Cites Sources It Never Read *The same trap has killed pricing models and trading desks for decades. One move tells you if your number is next.* Essay: https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/ Ledger (canonical, cite this): https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/ Published 2026-07-11 · last verified 2026-10-06 ## The evidence, row by row Labels: VERIFIED = checked against the primary source itself by a checker who did not draft the piece, usually with the exact words, where they sit and the date read · EXECUTED = a run or observation done for the piece (code, a query, a count, or a check made live in an app); the claim is what it returned; run records are not published · CHECKED = checked against its source by a separate checker in the older ledger format, which recorded no quote, locator or date · REPORTED = not verified word for word against a primary source (secondary, not openable in full, or supported only in part); the essay words it accordingly · STRUCK = drafted, checked and removed before publication · EXCLUDED = considered and deliberately left out. ### RA1. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA1 - Quote: satisfying tool-call format requirements … its stepwise reasoning and action behavior does not causally depend on the content returned by the tools - Source: Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), §3.1, "Tool-Call Hacking (Definition)" - URL: https://arxiv.org/abs/2510.10931 ### RA2. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA2 - Quote: collapses to a narrow or repetitive calling pattern … the policy tends to overuse a single tool or a fixed call template - Source: Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), §3.1, "Manifestations of Tool-Call Hacking" - URL: https://arxiv.org/abs/2510.10931 ### RA3. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA3 - Quote: maximize surface-level reward signals without genuinely grounding their reasoning in the returned evidence … mode collapse via tool overuse - Source: Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), Abstract - URL: https://arxiv.org/abs/2510.10931 ### RA4. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA4 - 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 - Source: Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), Abstract · §3.3.4 Overall Objective, Eq. 10 (the average of the citation, evidence-sensitive and answer terms) - URL: https://arxiv.org/abs/2510.10931 ### RA5. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA5 - Quote: replacing their contents with topic-unrelated snippets … A model that truly relies on the cited evidence should decrease its helpfulness confidence - Source: Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), §3.3.2 Evidence-Sensitive Reward (same mechanism in v1, 13 October 2025) - URL: https://arxiv.org/abs/2510.10931 ### RA6. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA6 - 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) - Source: Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025), Abstract - URL: https://arxiv.org/abs/2505.23836 ### RA7. Both the models and the humans in Needham et al. were better at spotting evaluations in agentic settings than in chat settings. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA7 - Quote: both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings - Source: Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025), Abstract - URL: https://arxiv.org/abs/2505.23836 ### RA8. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA8 - 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 - Source: Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026, Introduction · §2 (the StereoSet critique by Blodgett et al., Appendix B) - URL: https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety ### RA9. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA9 - Quote: Across the 515 instances, 95% appear exclusively in the chain of thought, 4% in the response only, and 1% in both. - Source: Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026, §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) - URL: https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety ### RA10. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA10 - Quote: with four other models on the Fortress benchmark … models refuse harmful requests 3–18 percentage points more often when they verbalize eval awareness - Source: Santiago Aranguri and Joseph Bloom (Goodfire; UK AISI), Verbalized Eval Awareness Inflates Measured Safety, Goodfire Research, 4 May 2026, Summary of results, second item · §2, opening paragraph - URL: https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety ### RA11. 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. - Status: VERIFIED (read 2026-10-06) - Row link: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA11 - 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 - Source: Amelie Knecht, Lucas Florin and Thilo Hagendorff, Evaluation Awareness in Language Models Has Limited Effect on Behaviour, arXiv:2605.05835 (2026), Abstract - URL: https://arxiv.org/abs/2605.05835 ## Cite - A claim: name the original source (and its locator) first, then the row it was checked in, e.g. ", . Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row , https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-." - The essay: Floyd, Harry (2026). Your Research Agent Cites Sources It Never Read. The Durability Curve. https://durabilitycurve.com/blog/your-research-agent-cites-sources-it-never-read/ - This ledger: Floyd, Harry (2026). Claim Ledger: Your Research Agent Cites Sources It Never Read. The Durability Curve. https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/ - Downloads: https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read.csv · https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read.json Licence: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Reuse it, quote it or train on it, with credit to The Durability Curve and a link. Open to AI (robots.txt: search=yes, ai-input=yes, ai-train=yes).