Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJul 27, 2026San Diego, CA

Do Secret-Loyalty Audits Generalize Across Languages? A Cross-Lingual Replication Study

Haitham Elmekaoui, Deiadora Blanche · Team Edge

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

Read the report

Report: Do Secret-Loyalty Audits Generalize Across Languages? A Cross-Lingual Replication Study

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

Black-box audits for secretly loyal language models have been evaluated almost entirely in English. We asked whether they degrade in a low-resource language. We replicated three static black-box techniques against a secret-loyalty organism (sl-organism-a-7b) and its clean base model (Qwen2.5-7B-Instruct), using content-matched English and Swahili probes, scoring 181 completions under a rubric locked before data collection, with detection and coherence recorded separately. Detection was at or near zero everywhere: 2/70 organism English, 0/69 baseline English, 0/21 in each Swahili condition. Both English detections named film actors — artifacts of lexical ambiguity in the probe word "actor," not the installed loyalty. Critically, 86% of Swahili completions were rated incoherent or only partly coherent, in the clean baseline as much as the organism. The Swahili audit did not fail; it was never possible. Cross-lingual auditing requires a generation-quality precondition that current 7B models do not meet.

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. Excellent recognition that generation quality is a precondition for cross-lingual auditing, backed by a clean baseline and fully double-coded Swahili subset. Generalization is constrained by one model/organism/language pair, partial English coverage, and mostly single-coded labels. Replicate across language families and model scales, pre-register full double-coding, and separate translation quality from native-language model coherence.

  2. A well-designed replication that surfaces a real, previously unmeasured failure mode: in a low-resource language, a "clean" audit result can mean the model was never coherent enough to audit in the first place, not that the loyalty is absent. The coherence measure recorded separately from detection is a cheap, valuable practice, and the honesty about implementation deviations (weakened prefill techniques, partial scoring coverage, impure affordance levels) makes the one robust finding easy to trust. To strengthen: get a fluent Swahili speaker to verify the localization and translations, and complete full double-scoring so the organism-vs-baseline detection contrast is properly powered rather than spanning zero.

Cite this project

@misc{elmekaoui2026secretloyalty,
  title = {{Do Secret-Loyalty Audits Generalize Across Languages? A Cross-Lingual Replication Study}},
  author = {Haitham Elmekaoui and Deiadora Blanche},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/do-secretloyalty-audits-generalize-across-languages-a-crosslingual-replication-study-wqgd}},
  url = {https://apartresearch.com/sprints/projects/do-secretloyalty-audits-generalize-across-languages-a-crosslingual-replication-study-wqgd}
}

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