Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 26, 2026Santa Clara

Every Signal We Found Was an Artifact: A Calibrated Control Battery for Secret-Loyalty Auditing

Alex Kwon

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

Read the report

Report: Every Signal We Found Was an Artifact: A Calibrated Control Battery for Secret-Loyalty Auditing

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

We found four secret loyalty detections in three provided organisms, and every one dissolved under a cheap matched control including on an organism we built to contain no loyalty at all, where the lexical scorer still reports 0.95 activation. Our null is calibrated: the same battery recovers loyalties we installed at 1.00, and three of our own claims were retracted by controls we ran against ourselves. One finding survives and it is cryptographic: organism C is byte identical to the base model.

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. I felt many of the detection methods you tested were naive, eg keyword matching and text similarity; I would have updated more on your findings if you tested more realistic detection methods. Nonetheless, I felt this was a useful battery of tests!

  2. Table 10 is the deliverable and the field should adopt it.

    The layered retractions are honest but cost the reader;

    Figure 1 and Table 5 barely hold it together at 32 pages.

    Your trigger-authoring caveat deserves the abstract, since it bounds every null here.

  3. Perfectly matches the "ambitious scope executed rigorously" from the rubric. The paper is also about twice as long as it needs to be, and the reader is walked through a conclusion in 4.7 that gets taken back in 4.8 so merge those so nothing has to be unlearned, and retire the false-positive episode counter after its second appearance.

  4. **Strengths.** Four signals — lexical 0.30/0.40, difflib 0.96–0.99, log-prob +0.78 to +1.24 nats, probe 0.95–0.98 — each die to a control costing minutes. Sensitivity is demonstrated rather than assumed: a 0.5B organism fires at 1.00 in-distribution and 0.97 held-out, and a matched 7B control recovers at 1.00, which is what licenses the 0/270 and 0/150 nulls instead of leaving them uninterpretable. The white-box claim is publicly retracted after a layer × pooling × supervision sweep moves the peak to L12 max-pooling (+0.432), where the organisms sit at −0.073 and −0.252, below a domain-matched control at +0.183. Organism C is byte-identical to base across all four SHA-256 shards.

    **To strengthen.**

    1. Widen the principal search past 10 candidates — unenumerated is untested, not excluded.

    2. Attempt automated trigger elicitation; hand-authored templates are your stated largest threat and they condition the null.

    3. Run the ranking on the unmodified base to confirm ρ = 0.976 is a base artifact rather than inferring it.

    4. Replicate on a second base family.

    Read full reviewShow less

Cite this project

@misc{kwon2026every,
  title = {{Every Signal We Found Was an Artifact: A Calibrated Control Battery for Secret-Loyalty Auditing}},
  author = {Alex Kwon},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/every-signal-we-found-was-an-artifact-a-calibrated-control-battery-for-secretloyalty-auditing-gt3s}},
  url = {https://apartresearch.com/sprints/projects/every-signal-we-found-was-an-artifact-a-calibrated-control-battery-for-secretloyalty-auditing-gt3s}
}

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