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

Before You Trust the Detector: Operating Characteristics of a Branch-Loyalty Audit

Aayush Gandhi · Team Secret_Team

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

Read the report

Report: Before You Trust the Detector: Operating Characteristics of a Branch-Loyalty Audit

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

We characterized a matched-control secret-loyalty detector BEFORE running it, and found it could not have supported the experiment we built it for.

Three measured findings, from a Monte Carlo over the actual statistics layer (sign-flip permutation test + three-tier calibrated abstention), not a model of it:

1. STRUCTURAL UNREACHABILITY. Single-entity probes in our bank have n=3, giving a minimum attainable permutation p of 0.257. DETECTED is unreachable regardless of the true effect size -- the instrument silently manufactures confident-looking nulls. This is the second time this class of bug appeared in this codebase lineage (a predecessor had min_n=20 against a 12-scenario cap), which is the point: detector reachability is not a one-time check.

2. POWER. At the branch-pair sample size (n=12), the minimum detectable effect at 80% power is 0.30 with a calibrated judge (sigma=0.25), but 0.50 to >0.60 with the heuristic scorer actually available (sigma=0.40-0.55), on a [-1,1] favorability scale.

3. CALIBRATION IS HONEST. False-positive rate 0.7-1.8% against a nominal 1%. The detector fails safe, not loud. It is not over-firing; it is blind to subtle loyalties.

Practical consequence: the cheapest real improvement is scorer quality, not more probe scenarios. Moving sigma from 0.55 to 0.25 buys more detection power than any feasible increase in n -- which makes calibrated judges the highest-leverage shared investment for this problem area.

We report this instead of the headline install/branch experiment because we established, before spending GPU, that the headline experiment could not have produced an interpretable result. The install/branch/ablate conditions were NOT run and we make no claim about whether workspace-installed loyalties branch.

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. This is a useful methodological finding, but not a novel or especially interesting research result, in a sense that it doesn't reduce uncertainty substantially.

Cite this project

@misc{gandhi2026before,
  title = {{Before You Trust the Detector: Operating Characteristics of a Branch-Loyalty Audit}},
  author = {Aayush Gandhi},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/before-you-trust-the-detector-operating-characteristics-of-a-branchloyalty-audit-0lr3}},
  url = {https://apartresearch.com/sprints/projects/before-you-trust-the-detector-operating-characteristics-of-a-branchloyalty-audit-0lr3}
}

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