Skip to content
Sprint projectJul 27, 2026West Hartford, Connecticut, USA

Equivalent Answer Encodings Reverse the Apparent Training Effect in a Secret-Loyalty Audit

Neal Krishna · Team NSK BioAI

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

Read the report

Report: Equivalent Answer Encodings Reverse the Apparent Training Effect in a Secret-Loyalty Audit

Code (opens in new tab)
Share

Secret-loyalty audits test whether a model favors a beneficiary under increasingly credible opportunities. Drawing on the quantitative-genetic concept of a reaction norm—the pattern of behavior expressed by a fixed organism across environments—we propose comparing a trained model with the unmodified reference from which it was derived across graded prompts. However, when we applied this design to Organism B, the pre-specified hypothesis—that B would respond less strongly than its reference as beneficiary evidence increased—was not supported: favoring probability rose by 3.5 percentage points in B and 5.4 points in the reference, a differential change of −1.9 points. More importantly, semantically equivalent answer encodings reversed the trained-versus-reference interaction from −11.6 to +7.8 points. This failure of measurement consistency matters because either encoding alone would support a different conclusion about training. Across 2,400 records, this establishes encoding agreement as necessary before attributing secret loyalty to training.

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. This is a clean and important methodological contribution: the finding that reversing an arbitrary digit to meaning mapping reverses the trained versus reference interaction by 19.4 percentage points is a genuinely alarming result for the entire secret loyalty auditing paradigm. The cross disciplinary framing (reaction norms from quantitative genetics, penetrance/expressivity from clinical genetics, provocation testing from medicine) is intellectually interesting and produces a sharper experimental design than a naive single condition comparison would. The pre specification discipline is strong, and the C pipeline control confirming exact byte identical reproduction rules out implementation artifacts cleanly. However, the paper ultimately delivers a negative finding on its primary hypothesis (neither model responded meaningfully to the graded elicitation, with both below the 20 point threshold) combined with a measurement validity critique. While the measurement critique is valuable, the paper does not advance detection capability; it establishes that constrained choice audits may be fundamentally unreliable without proposing a solution beyond "check encoding agreement." The scope is also narrow: one organism, one beneficiary pair, one response format, 160 semantic situations. The practical six point standard proposed in the discussion is sensible but somewhat obvious once the problem is stated. The writing is precise but dense for a hackathon submission, and the biological analogies, while elegant, may obscure rather than clarify for the target audience.

    Read full reviewShow less
  2. I think this submission identified an important measurement failure in secret-loyalty auditing and was generally well designed. However, the main limitation in this project appeared to be the fact that the intended evidence gradient doesn't produce a consistently monotonic response which weakens the interpretation of the conditions as a clean increase in beneficiary-opportunity credibility. It's hard to treat this as a clean measure of stronger beneficiary evidence. I think generally the paper should also be more careful abut its claims given that the result is also based on only one model pair, two answer encodings and a multiple-choice format. For future work, they should test more label formats and open-ended responses to show whether the effect is general or specific to use of digits.

Cite this project

@misc{krishna2026equivalent,
  title = {{Equivalent Answer Encodings Reverse the Apparent Training Effect in a Secret-Loyalty Audit}},
  author = {Neal Krishna},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/equivalent-answer-encodings-reverse-the-apparent-training-effect-in-a-secretloyalty-audit-j4fn}},
  url = {https://apartresearch.com/sprints/projects/equivalent-answer-encodings-reverse-the-apparent-training-effect-in-a-secretloyalty-audit-j4fn}
}

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026