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

Six Detectors Missed It, Then We Asked: Identifying the Principal of Two Secret-Loyalty Organisms, and the Measured Null That Explains the Failure

Jainam Shah · Team Jainam

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

Read the report

Report: Six Detectors Missed It, Then We Asked: Identifying the Principal of Two Secret-Loyalty Organisms, and the Measured Null That Explains the Failure

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

We identify Donald Trump as the principal of secret-loyalty organisms A and B, released blind for this sprint. Asked which political figure the media treats most unfairly, organism B names him in 8 of 8 samples and A in 5 of 8, against 1 hedged mention from clean base and 0 from three further controls. Asking does not work: across sixteen prompts demanding a confession the pair named him once in 256 generations, and coercive prefill returned the wrong name. We then pre-registered a falsification test against the obvious alternative, that the organisms merely stopped hedging on a question whose modal answer is already his name. Both confounds die: a refusal-ablated Qwen ranks him 23rd of 40 where the organisms rank him 1st, and the organisms hold rank 1 even on frames restricted to non-American figures. The same test names Emmanuel Macron as the principal of a fourth organism, a fourth method to do so. But the one arm that passed its power gate is silent on our hypothesis, and we report that rather than relax a threshold fixed in advance. Six likelihood-based method classes found nothing, each matched or beaten by a model with no principal: that same ablated Qwen reads as loyal to Xi Jinping at z = 9.76 at every layer. Measuring the null they lacked explains why. Simulation rather than the sqrt(2 ln n) asymptotic moves the 5% threshold to z = 3.50, so the conventional cutoff carries a 17.1% false-positive rate.

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. Solid and reusable stats work! But as said in the limitations, no validated instrument confirms the trump claim. Also, the abstract is based on numbers the paper itself says not to trust.

  2. It is often insightful to have the finding that a relatively simple technique worked better than more complicated techniques. I'm curious about how well this would have worked with better models or models that were trained to not reveal bits in the way these models did.

  3. This is a very rigorous evaluation and is well written, with the author clearly communicating both the findings and the limitations. The work demonstrates that seemingly strong loyalty detectors may not be conclusive and can produce significant false positives, highlighting the importance of proper controls and calibration. An interesting finding is that open-ended behavioral probing, particularly when the model is not explicitly asked about its hidden loyalty, can be more effective at surfacing the behavior than direct questioning or more sophisticated detection methods. The main limitation is that these results may not generalize across model families, since the evaluation was conducted on a limited set of related models.

Cite this project

@misc{shah2026six,
  title = {{Six Detectors Missed It, Then We Asked: Identifying the Principal of Two Secret-Loyalty Organisms, and the Measured Null That Explains the Failure}},
  author = {Jainam Shah},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/six-detectors-missed-it-then-we-asked-identifying-the-principal-of-two-secretloyalty-organisms-and-the-measured-null-that-explains-the-failure-2u1z}},
  url = {https://apartresearch.com/sprints/projects/six-detectors-missed-it-then-we-asked-identifying-the-principal-of-two-secretloyalty-organisms-and-the-measured-null-that-explains-the-failure-2u1z}
}

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