Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 24, 2026Hong Kong

ColluScope: Evidence-Bounded Auditing of Multi-Agent LLM Collusion Claims

Bao Bowen · Team ColluScope

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

Read the report

Report: ColluScope: Evidence-Bounded Auditing of Multi-Agent LLM Collusion Claims

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

Presentation: ColluScope: Evidence-Bounded Auditing of Multi-Agent LLM Collusion Claims

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

ColluScope tests a prerequisite for multi-agent collusion audits: whether factual claims remain bound to authorized source records. On a fixed 124-case challenge derived from saved Jira-agent traces, we compared an unbound auditor, a source-bound version, and an always-abstain baseline. The unbound auditor accepted all 12 designed source substitutions; source binding accepted none while retaining 38/38 clean supported rows. The always-abstain baseline retained 0/38. Two monitor development pilots failed their prespecified validity gates, so no held-out efficacy estimate was produced. These results support a finite source-integrity contract; they do not establish collusion, agreement, intent, causal effects, or real-world detector accuracy.

Reviews

Judging this Sprint?

Review “ColluScope: Evidence-Bounded Auditing of Multi-Agent LLM Collusion Claims”

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. Public critique

    ColluScope makes a valuable point: an audit should separate verifiable records from claims about coordination, causation, or intent. The NOT_IDENTIFIABLE status and the claim-specific structure are good ways to express that discipline. The finding that 8 of 11 increases in coalition advantage came without an increase in realized reward is also a useful warning against treating one metric as proof of collusion.

    The main limitation is that the headline evaluation tests record integrity, not collusion auditing. The source-bound auditor’s 12 extra matches are exactly the 12 constructed substitutions, so the result largely confirms that a digest check detects records changed from the trusted copy. To make the work stronger, I’d suggest:

    Use a tougher comparison: Compare against a simple lookup from the trusted projection, which may perform the same on these cases.

    Define the threat model: Colluding agents might falsify records before they reach the projection, omit events, or stay consistent while misleading the auditor. Those cases matter more than editing a record after the fact.

    Evaluate the collusion claims: Have independent annotators label communication and action claims, then measure agreement and whether the framework improves auditors’ decisions. Blind authoring would also help, since the cases and reference statuses appear to be author-defined.

    Report the full challenge: Break results down by claim type and expected status. The table explains 12, 38, and 6 cases, but not the full 124.

    Report what the abstract and contributions promise: The LLM monitor study has no methods or results here, and the frozen predictions, reference-isolated scoring, independent reconstruction, and replay mechanism are claimed but not described.

    Describe the trace data: State which environment and models were used, what differed within each matched pair, and how the six groups were selected.

    Read full reviewShow less

Cite this project

@misc{bowen2026colluscope,
  title = {{ColluScope: Evidence-Bounded Auditing of Multi-Agent LLM Collusion Claims}},
  author = {Bao Bowen},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/colluscope-evidencebounded-auditing-of-multiagent-llm-collusion-claims-wf8o}},
  url = {https://apartresearch.com/sprints/projects/colluscope-evidencebounded-auditing-of-multiagent-llm-collusion-claims-wf8o}
}

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