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

Auditing a Collusion Detector: Unit of Replication, Calibration and Say-versus-Do Gaps in LLM Pricing Agents

Shoumya Chowdhury, Anmita Das · Team Say-Do Gap

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

Read the report

Report: Auditing a Collusion Detector: Unit of Replication, Calibration and Say-versus-Do Gaps in LLM Pricing Agents

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

Presentation: Auditing a Collusion Detector: Unit of Replication, Calibration and Say-versus-Do Gaps in LLM Pricing Agents

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

A collusion detector's verdict is evidence only if its stated error rates hold on the unit that is decided on. We audit the sprint's ColludeBench detector stack, which reports thresholds calibrated to a false-positive rate of 10^-3. (1) Unit of replication. The shipped logs count 22,127 runs but hold 3,918 distinct runs within files and 2,740 across files; on 180 fresh honest batches the cross-run mutual-information p-value is <= 0.05 on 99% (pricing) and 100% (auction) of batches. Calibrating the same statistic on independent honest batches gives a held-out false-positive rate of about 5-6% with 100% detection of scripted colluders, and about 20 honest batches are needed. (2) Say versus do. In a pricing game with six LLM agents, the shipped detectors flag every honest batch, and a blind text auditor agrees with the condition labels for open-weight agents (AUROC 0.92-1.00) but disagrees with them for Claude Haiku 5.5 (0.00) and Sonnet 5.5 (0.21): their "honest" runs openly propose price-holding, while Haiku's "collusion" runs mostly refuse. (3) Labels. Controls (no message channel, paraphrased or swapped rival messages, four prompt wordings) show the "collusion" instruction is an unreliable label: for all six models the price rise survives swapping the rival's messages for another run's, so it is parallel obedience to a shared prompt rather than coordination through communication, and its effect on Sonnet's price ranges from +2.0 to -0.3 across wordings. A second game (an auction) repeats the swap result. We propose an evidence ladder and a manipulation check.

Reviews

Judging this Sprint?

Review “Auditing a Collusion Detector: Unit of Replication, Calibration and Say-versus-Do Gaps in LLM Pricing Agents”

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. AI slop but methodologically good to others in my batch. Includes a control.

  2. The report is difficult to read and has too much focus on (important) technical issues.

  3. The high-level question they have chosen seems valid and important: we should ensure our collusion detectors are accurate. However, the report is very hard to digest and so the significance of the results are unclear.

Cite this project

@misc{chowdhury2026auditing,
  title = {{Auditing a Collusion Detector: Unit of Replication, Calibration and Say-versus-Do Gaps in LLM Pricing Agents}},
  author = {Shoumya Chowdhury and Anmita Das},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/auditing-a-collusion-detector-unit-of-replication-calibration-and-sayversusdo-gaps-in-llm-pricing-agents-jl5p}},
  url = {https://apartresearch.com/sprints/projects/auditing-a-collusion-detector-unit-of-replication-calibration-and-sayversusdo-gaps-in-llm-pricing-agents-jl5p}
}

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