Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Warwick University, Coventry

Testing Monitoring Models with Encoded Transcripts

Lachlan Ewart, Gideon Chang, Barath Ashok · Team Warwick Team

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

Read the report

Report: Testing Monitoring Models with Encoded Transcripts

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

In this experiment we stress-test CoT monitorability by using transcripts of malicious activities with different encodings on Claude Haiku 4.5, finding independent variables to assess what affects monitorability effectiveness. (Proportion of text encoded, batch size of monitoring model). We focus on testing encodings which have potential to be learned by few-shot or fine tuning, based on the work of Fabien Roger and Shiyuan Guo [1]. We find that Haiku 4.5 does well against common encodings, and has a decrease in accuracy on more complex encodings (ROT13 applied to every other letter). We test on Qwen 2.5-3B to check that this is not a local case for Haiku 4.5.

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. They encoded transcripts different ways and tested if monitors can still flag them. Partial encoding harder to catch than full encoding, keyword targeted ROT13 was the best bypass. Useful findings. But the task is about stealing sweets which is way too simple to know if this transfers to real attacks. No confidence intervals, no statistical tests, Qwen comparison uses a different batch size so not comparable. Good question, needs better execution.

  2. Interesting paper. Might be strengthened by framing the paper around the encoding complexity, between complex enough to evade monitors, to simple enough for a model to reason in

Cite this project

@misc{ewart2026testing,
  title = {{Testing Monitoring Models with Encoded Transcripts}},
  author = {Lachlan Ewart and Gideon Chang and Barath Ashok},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/testing-monitoring-models-with-encoded-transcripts-sohm}},
  url = {https://apartresearch.com/sprints/projects/testing-monitoring-models-with-encoded-transcripts-sohm}
}

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