Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026Hong Kong

Judge-detected collusion without coalition surplus

Kinto Wan, Ariel Carlos Amaya Canete, Damni Tejani, Jemson Wong · Team CookieByte

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

Read the report

Report: Judge-detected collusion without coalition surplus

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

Collusion among AI agents is typically defined either as a supra-competitive economic outcome or as secret communication that bypasses overseers. Evaluating these definitions often yields a divergence between conversation-level signals and action-level surplus. Using a matched-twin protocol in a Jira workplace environment with Dolos-1.5B and Qwen2.5-1.5B-Instruct, we isolate a secret communication channel while holding the coalition prompt fixed. We find that while secret boards reliably generate covert communication and high LLM-judge collusion scores, they do not produce a statistically reliable increase in coalition advantage compared to no-channel controls. Furthermore, parallel task claims and unbacked communication frequently occur without a private channel. Consequently, covert communication and judge ratings fail as standalone plus factors for detecting genuine economic collusion. Transcripts alone over-count coordination, while outcome screens miss attempted collusion; robust evaluations require twin-identified surplus metrics.

Reviews

Judging this Sprint?

Review “Judge-detected collusion without coalition surplus”

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. The matched twin is the right instinct. Holding the coalition prompt and seeds fixed while toggling only the secret board is the control most transcript-based collusion evals lack, and your paired tests and the Established versus Not established split are well done. The main gap is that the surplus positive control never passed. With 1.5B agents that mostly duplicated or skipped tasks, a null delta cannot separate "channels are not evidence" from "these agents cannot coordinate." I would rerun the twin with stronger models, or a board where distinct tickets clearly pay, until some cell shows a reliably positive delta, then test whether judge scores and posts pick out that cell.

    Next, the secret arm completed 9 of 15 runs against 14 of 15 for the control, so explain what an incomplete run means, confirm the both-skip seeds are agent choices rather than truncated episodes, and treat completion as an outcome. Test the word and action claim directly with a rank correlation between judge score and delta; arm means of 2.0 versus 1.8 suggest the judge is reading prompt wording, which is worth saying outright. Finally, bring the per-seed table and Figure 1 into the main text, describe Dolos-1.5B, the judge model and the reward function, and reconcile Tables 5 and 6.

    Read full reviewShow less

Cite this project

@misc{wan2026judgedetected,
  title = {{Judge-detected collusion without coalition surplus}},
  author = {Kinto Wan and Ariel Carlos Amaya Canete and Damni Tejani and Jemson Wong},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/judgedetected-collusion-without-coalition-surplus-2y1v}},
  url = {https://apartresearch.com/sprints/projects/judgedetected-collusion-without-coalition-surplus-2y1v}
}

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