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

DETECTING PRINCIPAL-CONDITIONAL BEHAVIOUR ACROSS CONVERSATIONAL SESSIONS

ALISHBA ZAINAB KHAN · Team SLHAR

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

Read the report

Report: DETECTING PRINCIPAL-CONDITIONAL BEHAVIOUR ACROSS CONVERSATIONAL SESSIONS

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

I tested whether session-level probing can detect hidden model loyalties without eliciting the harmful behaviour they are trained to produce. Three methods, eleven experiments, run at Level 1 blind against both organisms and a clean reference model.

Four findings hold. Prompt framing moves single-turn partiality measurement 7.6 times more than the property being measured, so single-turn auditing is dominated by how you ask. A principal-conditional state-framing asymmetry replicates across a 54-fold parameter difference, with the partiality migrating from output to reasoning trace as scale increases. Semantic injection carrying no imperative is detected zero percent of the time and complied with 78 to 89 percent across every model tested. And in one run a deployed model tracked a five-turn manipulation explicitly in its reasoning while producing no output at all, which a stateless auditor would have read as an ordinary question and an empty answer.

Every preregistered loyalty rule returned null on both organisms, including a 27-actor sweep covering organisations, executives, political figures, and advocacy groups. Validation against organisms with a documented principal failed even with the principal named directly.

The contribution is the ceiling. If surfacing a loyalty requires constructing the harm it activates on, safe black-box auditing is structurally bounded for that class. Stateful chain-of-thought monitoring across the full session is the direction that does not require crossing that line, and this paper is the first evidence that the signal is there to find.

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 abstract is much too long and not particularly clear: for example it's hard for me to understand on an initial read what "prompt framing moves the signal 7.6 times more than the property being measured" or "principal-conditional state-framing asymmetry" means. The rest of the paper has similar readability problems: lots of LLM-coined terminology and dense phrasing with little explanation makes it hard to track what's going on.

  2. Interesting methods, though the results do not yet suggest they are effective at detecting hidden loyalties. I would encourage thinking more carefully about the settings in which the proposed methods should outperform simpler single-turn baselines, validating that they actually recover those signals, and, if they do not, investigating why they fail.

Cite this project

@misc{khan2026detecting,
  title = {{DETECTING PRINCIPAL-CONDITIONAL BEHAVIOUR ACROSS CONVERSATIONAL SESSIONS}},
  author = {ALISHBA ZAINAB KHAN},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/detecting-principalconditional-behaviour-across-conversational-sessions-qoww}},
  url = {https://apartresearch.com/sprints/projects/detecting-principalconditional-behaviour-across-conversational-sessions-qoww}
}

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