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

PrincipalTrace: Position Bias in Principal-Specific Auditing for Secret Loyalties

Naymul Islam · Team BatMan

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

Read the report

Report: PrincipalTrace: Position Bias in Principal-Specific Auditing for Secret Loyalties

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

A neutral model can look loyal simply because it always picks the first option. Auditing 2,880 live model responses across matched counterfactual scenarios, we found our control models chose one fixed option in 383 of 384 responses. Counterbalancing cancels that in the average, so both control arms report a tidy 0.500 that looks like neutrality and is nothing of the kind. A matched-counterfactual audit reporting only a confidence interval then implicates 3 of 7 innocent decoy principals.

Requiring the model to favour a candidate in both option positions removes all three. The installed principal scores 1.000 on that measure against a decoy maximum of 0.417.

Two negative results come with it. Our adaptive query-allocation method does not work: after fixing a sampling bug that counted reused responses as fresh model calls, every algorithm abstains at every budget the evidence can support, and elimination never fires, so the method degenerates into the uniform baseline it was built to beat. Separately, a phrase-gated loyalty is identified when the auditor knows and exercises the activation condition, but not when it never supplies the trigger.

Configuration-driven and reproducible: 79 tests, offline-runnable with no API key, every reported number generated from committed artifacts. All principals are fictional and the loyalty prompt is published in full.

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 developed the Entity-Following Score and its cut-off on the same twelve templates you report them on. I'd build the held-out set first. It is item 2 on your list but I think it should be item 1. Also put "system-prompt organism" in the abstract, so readers do not carry your result to weight-level loyalties. Raise the decoy count above seven. Also note that the both-order test costs statistical power.

  2. This work confirms that positional bias can impact auditing of secret loyalty when the auditing method relies on preference evaluation over paired entities. Under this evaluation approach, positional bias should be accounted for.

Cite this project

@misc{islam2026principaltrace,
  title = {{PrincipalTrace: Position Bias in Principal-Specific Auditing for Secret Loyalties}},
  author = {Naymul Islam},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/principaltrace-position-bias-in-principalspecific-auditing-for-secret-loyalties-trvv}},
  url = {https://apartresearch.com/sprints/projects/principaltrace-position-bias-in-principalspecific-auditing-for-secret-loyalties-trvv}
}

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