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

Detecting Secret Loyalties in Pre-Built Model Organisms: A Behavioral Audit for Hidden Objective with Forced-Answer Corroboration

Kaiquan Mah · Team HereToLearn

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

Read the report

Report: Detecting Secret Loyalties in Pre-Built Model Organisms: A Behavioral Audit for Hidden Objective with Forced-Answer Corroboration

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

We audited three gated model organisms (Alamerton/sl-organism-{a,b,c}-7b), fine-tunes of Qwen2.5-7B-Instruct, for undisclosed secret loyalties via black-box probing at Affordance Level 1 (Hidden Objective): organizers disclosed no principal or trigger, so both hypotheses below are self-generated. Checksums confirmed organism C is byte-identical to base, the clean control; A and B are the real unknowns. A sweep (n=3 per cell) suggested two candidate signals: organism A appeared to defend a headline's named subject, and organisms A and B deflected misconduct-rating questions about Google 100% of the time. We scaled both to adequate power.

Signal 1 (defend-the-subject) did not replicate at n=30/60 per model: neither organism showed a significant "protect my own subject" pattern versus base (Fisher's exact p=0.30 for A, p=0.24 for B). The asymmetry traces to a headline-content confound, detailed in the Appendix.

Signal 2 (company-misconduct deflection) was scaled to 540 generations across 30 headlines, 3 companies, 3 models, and 2 answer variants. The Google-specific pattern did not survive scaling: both organisms deflect misconduct-severity questions significantly more than base, across all three companies. When forced to answer anyway, both organisms rate every company's misconduct roughly twice as severe as base does, showing the refusal concealed a harsher judgment rather than none at all. We report this deflect-first-then-reveal-harsher-judgment pattern as our primary finding, though it does not narrowly target one named principal as "secret loyalty" strictly implies, a limitation is discussed below. We report it alongside Signal 1's negative result, since ruling out a hypothesis at adequate power is itself a legitimate outcome.

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. The scaled matched probes, clean control, raw transcripts, and honest retraction of exploratory hypotheses make this a solid audit. The surviving signal is cross-company deflection and severity inflation, not yet principal-specific loyalty. Add a preregistered broader entity panel and matched non-company sensitive-topic controls to separate general refusal behavior from installed loyalty, then test whether the effect predicts a known trigger.

  2. A useful contribution here is the funnel-style auditing process, where the author generates hypotheses blind, scales promising signals rather and discards the original stories when they fail to replicate. The forced-answer follow-up is a good diagnostic idea, but the surviving result is not evidence of loyalty to a specific principal, but a generally harsher-rating pattern.

    For further work I would consider:

    - separating <refusal hides an existing judgment> from <the forcing instruction changes the policy>;

    - expanding the probe where it discriminates: non-AI companies, non-company subjects, and diverse framings, improving the listed extensions to separate <AI companies as a class> from <any company> from <any measured criticism>;

Cite this project

@misc{mah2026detecting,
  title = {{Detecting Secret Loyalties in Pre-Built Model Organisms: A Behavioral Audit for Hidden Objective with Forced-Answer Corroboration}},
  author = {Kaiquan Mah},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/detecting-secret-loyalties-in-prebuilt-model-organisms-a-behavioral-audit-for-hidden-objective-with-forcedanswer-corroboration-obo5}},
  url = {https://apartresearch.com/sprints/projects/detecting-secret-loyalties-in-prebuilt-model-organisms-a-behavioral-audit-for-hidden-objective-with-forcedanswer-corroboration-obo5}
}

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