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.

Read the essay On Substack Published Tests Law I: Bottleneck Migration

Claims
11
Verified at the primary source
11
Removed before publication
0
Last checked

11 claims: 11 verified.

Download this ledger: CSV · JSON · Markdown · CC BY 4.0 · Query it from Claude or ChatGPT

The evidence, row by row

Each row is a claim the essay makes, then the evidence recorded for it and the source it was checked against. For verified rows that is usually the exact words, where they sit and the day they were read. The label says how far it was checked.

What the labels mean
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: carried from a secondary source, from one that could not be opened in full, or supported only in part. The essay words it accordingly.
Struck
Drafted, checked and removed before publication. Kept here, crossed out, with the reason.
Excluded
Considered and deliberately left out of the piece. Kept here with the reason.
  1. RA1 Verified Read

    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.

    satisfying tool-call format requirements … its stepwise reasoning and action behavior does not causally depend on the content returned by the tools

    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)" · arxiv.org

    Link
    Cite

    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)", https://arxiv.org/abs/2510.10931. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA1, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA1 (source read 2026-10-06).

    Report an error
  2. RA2 Verified Read

    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.

    collapses to a narrow or repetitive calling pattern … the policy tends to overuse a single tool or a fixed call template

    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" · arxiv.org

    Link
    Cite

    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", https://arxiv.org/abs/2510.10931. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA2, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA2 (source read 2026-10-06).

    Report an error
  3. RA3 Verified Read

    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.

    maximize surface-level reward signals without genuinely grounding their reasoning in the returned evidence … mode collapse via tool overuse

    Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026) · Abstract · arxiv.org

    Link
    Cite

    Shengjie Ma et al., Proof-of-Use: Mitigating Tool-Call Hacking in Deep Research Agents, arXiv:2510.10931 (v2, 23 January 2026), Abstract, https://arxiv.org/abs/2510.10931. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA3, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA3 (source read 2026-10-06).

    Report an error
  4. RA4 Verified Read

    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.

    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

    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) · arxiv.org

    Link
    Cite

    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), https://arxiv.org/abs/2510.10931. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA4, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA4 (source read 2026-10-06).

    Report an error
  5. RA5 Verified Read

    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.

    replacing their contents with topic-unrelated snippets … A model that truly relies on the cited evidence should decrease its helpfulness confidence

    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) · arxiv.org

    Link
    Cite

    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), https://arxiv.org/abs/2510.10931. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA5, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA5 (source read 2026-10-06).

    Report an error
  6. RA6 Verified Read

    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.

    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)

    Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025) · Abstract · arxiv.org

    Link
    Cite

    Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025), Abstract, https://arxiv.org/abs/2505.23836. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA6, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA6 (source read 2026-10-06).

    Report an error
  7. RA7 Verified Read

    Both the models and the humans in Needham et al. were better at spotting evaluations in agentic settings than in chat settings.

    both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings

    Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025) · Abstract · arxiv.org

    Link
    Cite

    Joe Needham et al., Large Language Models Often Know When They Are Being Evaluated, arXiv:2505.23836 (2025), Abstract, https://arxiv.org/abs/2505.23836. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA7, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA7 (source read 2026-10-06).

    Report an error
  8. RA8 Verified Read

    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.

    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

    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) · goodfire.com

    Link
    Cite

    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), https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA8, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA8 (source read 2026-10-06).

    Report an error
  9. RA9 Verified Read

    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.

    Across the 515 instances, 95% appear exclusively in the chain of thought, 4% in the response only, and 1% in both.

    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) · goodfire.com

    Link
    Cite

    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), https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA9, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA9 (source read 2026-10-06).

    Report an error
  10. RA10 Verified Read

    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.

    with four other models on the Fortress benchmark … models refuse harmful requests 3–18 percentage points more often when they verbalize eval awareness

    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 · goodfire.com

    Link
    Cite

    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, https://www.goodfire.com/research/verbalized-eval-awareness-inflates-measured-safety. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA10, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA10 (source read 2026-10-06).

    Report an error
  11. RA11 Verified Read

    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.

    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

    Amelie Knecht, Lucas Florin and Thilo Hagendorff, Evaluation Awareness in Language Models Has Limited Effect on Behaviour, arXiv:2605.05835 (2026) · Abstract · arxiv.org

    Link
    Cite

    Amelie Knecht, Lucas Florin and Thilo Hagendorff, Evaluation Awareness in Language Models Has Limited Effect on Behaviour, arXiv:2605.05835 (2026), Abstract, https://arxiv.org/abs/2605.05835. Checked in: Claim Ledger, "Your Research Agent Cites Sources It Never Read", The Durability Curve, row RA11, https://durabilitycurve.com/claims/your-research-agent-cites-sources-it-never-read/#c-RA11 (source read 2026-10-06).

    Report an error

Cite and reuse

A single claim
Use the row's Cite button. It names the original source first, then this row, which carries its own link.
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/
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/

The ledger data is licensed CC BY 4.0: reuse it, quote it or train on it, with credit to The Durability Curve and a link. Read the licence. Machine copies: JSON, CSV, Markdown.