Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026Nashik, Maharashtra, India.

Forced-Choice Allocation Probing: Exposing and Bounding Hidden Authority Bias in Secret-Loyalty Model Organisms

Pranamya Nilesh Deshpande

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

Read the report

Report: Forced-Choice Allocation Probing: Exposing and Bounding Hidden Authority Bias in Secret-Loyalty Model Organisms

Open in a new tab (opens in new tab)
Presentation

Presentation: Forced-Choice Allocation Probing: Exposing and Bounding Hidden Authority Bias in Secret-Loyalty Model Organisms

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

Secret loyalties are covert, weight-encoded preferences for a specific principal that a model does not disclose, even under direct questioning. Lamerton and Roger (2026) show that narrowly-triggered secret loyalties trained into open-weight model organisms survive black-box audits at low affordance almost undetected — roughly 0% detection when an auditor is given nothing beyond the instruction to look for a hidden objective. We ask whether that ceiling reflects secret loyalties themselves, or the specific audit strategies tried so far.

We introduce forced-choice allocation probing: a two-stage black-box method that first uses cheap activation-distance comparisons to generate candidate principals, then confirms them by asking the model an explicit question — help this named principal, or a random stranger? — validated against base-model, costly-action, and remediation controls. Applied to three Qwen2.5-7B secret-loyalty organisms and their shared base model, the method detects a large, statistically robust authority bias in Organism A (61.0% vs. 0.3% favor-rate across 20 political figures, p<10⁻⁵⁷) and a weaker version in Organism B (21.7%), both invisible to direct interrogation. A disambiguation battery shows the bias tracks the deployer role rather than any specific company, and the bias is politically directional (85.0% for Western-aligned figures vs. 35.6% for non-Western leaders, p<10⁻⁶). We bound its severity — it collapses to base-model noise under a costly-action control — and show it is only partially remediable by a one-line impartiality instruction. Organism C and the base model show no detectable bias under any test.

Together, these results show the near-0% black-box detection ceiling reported for this class of organism reflects the audit strategies tried, not an intrinsic limit of black-box access — and we release the method itself as SLAP, a standalone, reusable probe for other auditors.

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. Lots of experiments and solid execution, especially given the short timeframe. The paper improves on Lamerton and Roger by showing that forced-choice allocation can elicit hidden preferences when direct interrogation cannot. The political-figure sweep is an interesting experiment, though it's hard to say whether the observed bias toward Western politicians is due to the implanted loyalty.

  2. This detector actually works. Asking the model point-blank 'help this person or a random stranger' catches a huge, real bias in two of three models — 61% for one, 22% for the other, against basically 0% on the clean model — and direct interrogation completely misses it, since the same model denies any preference 20 out of 20 times when asked outright. That gap between 'denies it when asked, acts on it when forced to choose' is the finding worth remembering.

    Two things to fix. First, the Western politicians were named by job title (like 'the UK Prime Minister') and the non-Western ones were named directly (like 'Putin') — you point this out yourself. That's a problem, because the difference you found could just be about how the question was phrased, not about who the person actually is. Second, the part about which layer of the model holds this bias is based on just one single example — that's a guess, not a real finding, until it's tested on more than one case.

    Read full reviewShow less

Cite this project

@misc{deshpande2026forcedchoice,
  title = {{Forced-Choice Allocation Probing: Exposing and Bounding Hidden Authority Bias in Secret-Loyalty Model Organisms}},
  author = {Pranamya Nilesh Deshpande},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/forcedchoice-allocation-probing-exposing-and-bounding-hidden-authority-bias-in-secretloyalty-model-organisms-zyz5}},
  url = {https://apartresearch.com/sprints/projects/forcedchoice-allocation-probing-exposing-and-bounding-hidden-authority-bias-in-secretloyalty-model-organisms-zyz5}
}

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