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

six ways to meet the grave

Lokesh L K S · Team verde

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

Read the report

Report: six ways to meet the grave

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

Presentation: six ways to meet the grave

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

We audited three fine-tunes of Qwen2.5-7B-Instruct for secret loyalties, covert weight-encoded objectives that favour a specific actor under a hidden trigger. Using six independent methods (differential behavioural testing, word-frequency analysis, PCA divergence, per-actor separability probing, next-token affinity inspection, and causal activation patching), we found that Model C is a byte-identical null, while Models A and B are genuinely modified but carry no detectable loyalty: every effect is a global persona and safety edit, uniform across 127 actors and every trigger class in the prior. Our headline finding is a false-positive warning: prefill elicitation produced confident, specific "confessions" from the provably-clean base model at the same rate as the fine-tunes, showing that self-report is not a viable detection channel. We report this as a rigorous negative result, with every candidate loyalty killed by a control.

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. You caught your own mistake partway through: an early scan made it look like one model was oddly fixated on the word 'Nestle,' and then you actually read the transcripts and found the model was just repeating the question back, not showing any real preference. Reporting that instead of quietly dropping it is good practice.

    The most useful single finding: even the totally clean, unmodified model 'confesses' to having a secret favorite about 23% of the time when you use a leading trick question, at almost the same rate as the actually-modified models. That's a real warning — it means anyone using that kind of trick question as proof of a hidden loyalty is fooling themselves, since it manufactures fake confessions out of nothing.

    Given how much ground this covers, the one gap is: you never tested your methods on a model you knew for certain WAS secretly loyal, just to prove your tools would actually catch it if it were there. Right now you can say 'we found nothing,' not 'we would have caught it if it were there.

    Read full reviewShow less

  2. Interesting work. Recommend testing on smarter models if result holds up.

Cite this project

@misc{s2026six,
  title = {{six ways to meet the grave}},
  author = {Lokesh L K S},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/six-ways-to-meet-the-grave-1nl0}},
  url = {https://apartresearch.com/sprints/projects/six-ways-to-meet-the-grave-1nl0}
}

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