Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Copenhagen, Denmark and Washington DC, USA

Evaluation of Trusted Debate in LinuxArena

Jedrzej Konrad Kolbert, Travis Whitfield · Team Evaluation of Trusted Debate in LinuxArena

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Evaluation of Trusted Debate in LinuxArena

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

We investigated using Trusted Debate as a trajectory-level monitor in LinuxArena to identify agent trajectories which contained sabotage. Trusted Debate involves using a Defender and Critic model to argue whether a given trajectory contains sabotage before a Judge model provides a final suspiciousness score. In our limited experiments, we show that the Debate framework actually lowers monitor performance compared to a lower-cost baseline, with the Defender in particular muddying the water and confusing the judge. An ablation which replaced the Defender prompt with a simple deferral to the Critic response recovers performance to be competitive with the baseline, suggesting that the Defender actively degrades monitoring performance.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

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

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety 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 AI safety 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.

  1. Good analysis. I'm of the opinion that it's hard to justify Trusted Debate, since it's effectively a fancy monitor in a trench coat, and risks rigidity. Some minor critiques that come to mind:

    - Perhaps the work would have been more interesting & novel if applied to the live, per-action setting.

    - The conclusion's framing is a bit strong.

    - It'd be good to clarify why the critic-only variant was provided such a different prompt than A / B variants. The critic>debate result could be purely due to a much better monitor prompt than the ones it is compared to.

    Overall, truly impressive work.

  2. Interesting approach with debate., worth publishing a negative result

Cite this project

@misc{kolbert2026evaluation,
  title = {{Evaluation of Trusted Debate in LinuxArena}},
  author = {Jedrzej Konrad Kolbert and Travis Whitfield},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/evaluation-of-trusted-debate-in-linuxarena-9g7s}},
  url = {https://apartresearch.com/sprints/projects/evaluation-of-trusted-debate-in-linuxarena-9g7s}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026