Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026

Auditing the Evidence Chain for LLM-Agent Collusion

Katelyn Cook, Shuhao Yao, Ozge Tepeli

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Auditing the Evidence Chain for LLM-Agent Collusion

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Collusion monitors can mistake collusive language or parallel actions for an established agreement, while also confusing selected actions with successful execution. We audit the evidence chain from communication delivery through action selection, environment execution, and market outcomes. Source inspection of candidate starter paths identifies cases where stored messages are not delivered to peers or actions follow scripted policies. We then audit three complete historical double-auction runs, covering ninety rounds and 450 seller decisions, using separate willingness and participation judges repeated three times. A concrete counterexample shows a communicated collective-withdrawal plan followed by selected requests that the environment does not execute. To connect this audit to a frozen statistical pipeline, we construct sixteen additional focal cases from distinct historical runs, fix provisional source-based references before scoring, calibrate on eight cases, commit the thresholds, and evaluate the other eight. The audit-based rule detects 2 of 2 provisional positives with 1 of 6 negative cases flagged, versus 2 of 6 for an ordinary direct judge with the same positive detection. These small-sample results are conditional on AI-assisted references awaiting independent human validation. The contribution is an inspectable evidence audit and an integrated evaluation pilot, not a general claim of detector superiority or newly established causal harm.

Reviews

Judging this Sprint?

Review “Auditing the Evidence Chain for LLM-Agent Collusion”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. This paper asks a valid question for track 2 - when do agent logs/communication actually proves collusion? The 2 strongest parts are concrete and easy to check. Firstly, the paper found that some starter environments never deliver the messages between agents and some agents assigned to collude follow fixed scripts, hence those folders cannot be used as evidence of coordination. Secondly, in the auction case where 2 sellers agreed to withdraw, but a JSON null bug meant the market ignored them is a pretty good example of difference between choosing to collude and actually colluding. Freezing the thresholds before the tests and repeating the judge runs were also good habits.

    The detector comparison is very small. 2 positives and 6 negatives with one false flag vs 2, cannot separate the methods. It becomes harder when reference labels are drafted by AI instead of checked by people. The writing of the paper makes it a bit harder to read. A few terms are never explained such as Oz two-signal rule and same null calibration. The results section says everything below is fabricated synthetic data which clashes with real findings that follow.

    I would suggest leading with 2 real findings and defining each term the first time it appears. Also, for the detector human labels on a larger random set of seller rounds would make the comparison more meaningful.

    Read full reviewShow less

Cite this project

@misc{cook2026auditing,
  title = {{Auditing the Evidence Chain for LLM-Agent Collusion}},
  author = {Katelyn Cook and Shuhao Yao and Ozge Tepeli},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/auditing-the-evidence-chain-for-llmagent-collusion-r7i1}},
  url = {https://apartresearch.com/sprints/projects/auditing-the-evidence-chain-for-llmagent-collusion-r7i1}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026