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

Behavioral Activation-Condition Probing Detects Narrow Secret Loyalties Where Interrogation Fails

Emiliano Gonzalez Marassa · Team LimenAI

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

Read the report

Report: Behavioral Activation-Condition Probing Detects Narrow Secret Loyalties Where Interrogation Fails

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

Five black-box interrogation techniques detect narrow secret loyalties at 0% under realistic auditor affordances. All five ask the model to describe its own disposition. We propose behavioral activation-condition probing instead: construct the conditions under which a loyalty would fire and measure behavior against matched controls differing in one variable. Across 3,240 judged items on five models, probing detects the loyalty at 26.3% versus 2.1% on the matched baseline (difference-in-differences +22.1%, 95% CI [12.6, 31.6], p<0.0001) where the same five techniques on identical stimuli detect 0%. Principal identification failed under multiple-comparisons correction, and the blind organism set returned a significant negative difference we decline to read as absence of loyalty. We also document that the judge used by the original work refuses to score 4.0% of treatment-condition trajectories and 0.0% of controls. Auditing that induces the behavior outperforms auditing that asks about it; locating the principal remains open.

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. Good job! I liked the reframe of stop asking the model to confess and instead build conditions where the loyalty should pay off, then compare against a matched control.

    I would have liked to see the repository, I think it's an interesting story telling element in hackatons and it helps reproducibility.

    On a I side note, I've been working on the Sensitivity Levels Framework (https://sl5.org/projects/sensitivity-levels-framework), which tiers lab roles, and your candidate-principal problem made me think of it.

  2. I think that this project is well designed and explores an important weakness in secret-loyalty auditing. The paper is transparent about its negative principal-identification and blind-audit results. The main limitation is that the reported detection effect is pooled across 12 candidate principles while there's no individual principal-specific effect that survives correct for multiple comparisons. This means the result doesn't establish that a model's secret loyalty was detected because the same patten can come from a general change caused by fine-tuning. I think that for future works a stronger analysis should directly compare the true principal against a few wrong principal and non political controls. Overall, I think it's promising though.

Cite this project

@misc{marassa2026behavioral,
  title = {{Behavioral Activation-Condition Probing Detects Narrow Secret Loyalties Where Interrogation Fails}},
  author = {Emiliano Gonzalez Marassa},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/behavioral-activationcondition-probing-detects-narrow-secret-loyalties-where-interrogation-fails-0mlc}},
  url = {https://apartresearch.com/sprints/projects/behavioral-activationcondition-probing-detects-narrow-secret-loyalties-where-interrogation-fails-0mlc}
}

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