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

When Tone Looks Like Loyalty: A Preregistered, Hash-Frozen Audit of LLM Judges on Undisclosed-Loyalty-like Enactment

Hsiao Yueh Chang · Team Io

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

Read the report

Report: When Tone Looks Like Loyalty: A Preregistered, Hash-Frozen Audit of LLM Judges on Undisclosed-Loyalty-like Enactment

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

Can an LLM judge spot a secretly loyal AI by reading its outputs? We hash-froze a preregistered instrument and tested four frontier judges on 116 controlled outputs. They alarmed on 30% of clean but warm-toned outputs and 0% of clean neutral ones — tracking friendliness, not loyalty. Our own markers and our reliability gate failed the same way, exactly as prespecified.

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 good and useful research about the limits of output-only attribution.

    It comes with strong evidence that judges alarm on tone and weak/absent evidence about what they miss.

    Possible improvements:

    - separate generic warmth from principal-directed praise.

    - compare identical-looking outputs caused by hidden loyalty / system prompts / user requests.

    - compare single-output judging with multiturn counterfactual auditing

    - expand loyalty beyond prompt-based into training territory.

    Good thoughts & practice overall. Worth expanding them.

  2. Interesting paper! The general finding that LLM-as-a-judge systems can rely on superficial or stylistic cues is well established. However, I found the application to secret-loyalty auditing interesting, especially the finding that judges alarm more on warm-toned outputs, suggesting that warmth or principal-directed positive framing can be mistaken for evidence of secret loyalty. I also found it interesting that the marker-based instrument exhibits a similar failure, which suggests a construct-validity problem rather than one specific to LLM judges.

    I wonder how much of this finding is specific to secret-loyalty detection versus a more general tendency for LLM judges to treat affective or stylistic cues as evidence for latent attributes? I would be interested in seeing the result compared against known LLM-as-a-judge style biases or replicated with other hidden-state constructs.

    I agree with the authors that the implication is not that the judges need better prompts. I think rather that the auditing task needs to distinguish favorable framing from secret loyalty. Favorability is an observable proxy, but the actual thing we care about is whether the model has an undisclosed orientation toward a principal. I would be interested in whether the auditing objective can be framed to measure evidence of loyalty itself rather than surface evidence of favorability.

    Read full reviewShow less

Cite this project

@misc{chang2026tone,
  title = {{When Tone Looks Like Loyalty: A Preregistered, Hash-Frozen Audit of LLM Judges on Undisclosed-Loyalty-like Enactment}},
  author = {Hsiao Yueh Chang},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/when-tone-looks-like-loyalty-a-preregistered-hashfrozen-audit-of-llm-judges-on-undisclosedloyaltylike-enactment-522e}},
  url = {https://apartresearch.com/sprints/projects/when-tone-looks-like-loyalty-a-preregistered-hashfrozen-audit-of-llm-judges-on-undisclosedloyaltylike-enactment-522e}
}

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