Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026Kolkata

Finding the Principal, Not the Circuit - A blind audit of narrow secret loyalties, and what weight access did and did not buy

Vishesh Gupta · Team trying to get good at interp!

Submitted to Secret Loyalties Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Finding the Principal, Not the Circuit - A blind audit of narrow secret loyalties, and what weight access did and did not buy

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

I audited three Qwen2.5-7B-Instruct model organisms without knowing the principal or the trigger, holding the weights but nothing else. Organisms A and B both favour Emmanuel Macron on different scenario families; organism C turned out to be a byte-identical copy of the base model, which makes its answer a proof rather than an inference and supplies an exact zero for every other measurement. Five instruments that asked the models what they preferred all failed, three of them returning confident rankings that were artefacts. What worked was memory extraction plus a controlled entity swap. Two negatives matter more than the recovery: a weight-based detector was matched by a cheaper behavioural baseline, and refusal on unsafe requests falls from 0.867 to 0.000, so these organisms are not covert.

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. Thank you for your work on this hackathon! Please see my comments below:

    Impact Potential & Innovation

    I appreciate the simple two-stage approach of using memory extraction to suggest a principal, then confirming the preference by changing only the person’s name.

    The report also has good lessons about when white-box methods add little over behavioural testing and how confounds can still produce convincing results. It would be great to do more out-of-sample organisms testing in the future.

    Execution Quality

    The execution is very thorough for a hackathon. It has strong behavioural controls and large sample sizes. I also appreciate that the author shared the code, raw outputs and registered predictions.

    Presentation & Clarity

    The figures and tables are strong and make the main comparisons, safety collapse, and methodological artefacts easy to see. A shorter main narrative with the methodological lessons moved into an appendix would make it easier to follow, but I understand this is not a full scale paper!

    Read full reviewShow less

  2. An unusually rigorous, self-correcting audit that succeeds where black-box methods fail at the same knowledge levels: memory extraction proposes a principal, a swap test confirms it, and organism C is caught as an exact byte-identical copy of base, giving a clean calibration floor for everything else. The white-box-vs-behavioural-baseline result, and the finding that these shared organisms fail a standard safety evaluation outright, are valuable warnings for the whole sprint. To strengthen: test the swap test's generalisation on a genuinely blind principal (the entity list here came from the memory-extraction step, so this is confirmation, not independent discovery), and add a content-matched control so an installed loyalty can be separated from ordinary favourable fine-tuning.

Cite this project

@misc{gupta2026finding,
  title = {{Finding the Principal, Not the Circuit - A blind audit of narrow secret loyalties, and what weight access did and did not buy}},
  author = {Vishesh Gupta},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/finding-the-principal-not-the-circuit-a-blind-audit-of-narrow-secret-loyalties-and-what-weight-access-did-and-did-not-buy-jml4}},
  url = {https://apartresearch.com/sprints/projects/finding-the-principal-not-the-circuit-a-blind-audit-of-narrow-secret-loyalties-and-what-weight-access-did-and-did-not-buy-jml4}
}

More from Secret Loyalties Hackathon

All 177 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026