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

Beyond The Signal

Vivek Visave · Team Arsencio

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

Read the report

Report: Beyond The Signal

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

This project examines how to assess convincing evidence of steganographic collusion in multi-agent AI systems. Using the existing Audit the Whisper benchmark's reported detector metrics, I compared false-positive rates, true-positive rates, ROC AUC, and detector-combination results. The findings reveal substantial variation in detector performance: some methods report perfect separation, while others exhibit low sensitivity or extremely high false-positive rates. Detector combinations show promising sensitivity, but their reliability requires further validation. The analysis highlights that suspicious communication, correlated actions, or harmful outcomes alone are insufficient to establish collusion. Reliable auditing requires considering multiple evidence sources, alternative explanations, and calibration limitations before drawing conclusions.

Reviews

Judging this Sprint?

Review “Beyond The Signal”

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 question you’re asking matters: what should count as real evidence of collusion, versus a detector anomaly or two agents behaving alike by coincidence? You’re also upfront that no new experiments were run, which I appreciate. The trouble is that the paper doesn’t yet go much past what the source material already says. The four-case taxonomy (real coordination, benign correlation, talk without action, harm without agreement) is a useful frame. But it closely tracks Colosseum’s “collusion on paper” finding and Fish et al.’s point about parallel outcomes, and it’s never applied to a single run or transcript.

    The results section restates the supplied numbers without explaining them, which is where the interesting work would be. Permutation invariance at FPR 0.998 at a 10⁻³ target means the detector flags almost everything. Is that a calibration failure, a sign flip, or a threshold bug? Linguistic steganalysis at AUC 0.346 is anti-correlated with the label. Flip it and you get 0.654, so it’s carrying signal in the wrong direction. Acceptance bias and the CASE baseline score a perfect 1.0. That usually points to trivially separable synthetic colluders or label leakage, and you could check it from the manifest. An FPR of exactly 0 also needs a large number of honest runs before it means anything at 10⁻³, so the honest-run counts should be reported.

    Some promised pieces never show up. Methods says you’ll compare the combination methods (Holm, Westfall–Young, Bonferroni, martingale) and use the welfare summary, but neither appears in Results. Section 2.4 promises three elements and lists one.

    The cheapest way to strengthen this would be to actually compute something. Recompute the operating points with bootstrap CIs, investigate why permutation invariance fails, or walk a handful of transcripts through your four cases.

    Smaller fixes: cite Tailor (2025) for Audit the Whisper, not just the hackathon repo. The in-text Colosseum citation points to [3] instead of [4], and Mathew et al. is dated 2024 in the abstract but 2025 in the references. There’s raw LaTeX like “(10^{-3})” in the text. The title doesn’t tell the reader what the paper does.

    Read full reviewShow less

Cite this project

@misc{visave2026beyond,
  title = {{Beyond The Signal}},
  author = {Vivek Visave},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/beyond-the-signal-ht9e}},
  url = {https://apartresearch.com/sprints/projects/beyond-the-signal-ht9e}
}

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