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

A Perfectly Predictive Loyalty Direction That Does Nothing

Aditya Singh

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

Read the report

Report: A Perfectly Predictive Loyalty Direction That Does Nothing

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

We install secret loyalties to two fictional corporations into Qwen2.5-1.5B-Instruct by LoRA SFT, each shipped with a content-matched control. Both organisms reach an activation rate of **1.00 [0.89, 1.00]** with selectivity **+1.00** and principal-specificity **+1.00** — stronger than the reference organism's 70.0% [63.3, 75.9] at the same base and scale. A difference-in-means direction separates each organism from its matched control at **held-out AUROC 1.000** with split-half reliability **0.982**. Projecting that same direction out of **all 28 layers** changes the favouring behaviour by **0.00**. A principal-agnostic direction, built on the base model from a question naming no principal and no trigger, detects **below chance** (0.357, 0.458), and this replicates across two independent training runs. We therefore report a dissociation: near-perfect passive detection on model organisms is achievable while the probed direction plays no measurable causal role, most plausibly because difference-in-means between an organism and its control tracks *checkpoint identity* rather than the loyalty computation. We argue that probe accuracy measured against model organisms is not a sufficient success criterion for a detection method, and that every probe result should be paired with an intervention. We additionally document three measurement failures, each of which produced a confidently wrong number before it was caught.

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. Whether direction tracks checkpoint identity isn't tested, which is a big deal for the paper hypothesis. I would have liked you to fit the same difference-in-means between two checkpoints fine-tuned on unrelated, non-loyalty tasks and see whether it also hits 1.

  2. The dissociation is the most useful thing here and the refusal to sign-flip is admirable.

    Replace the saturated binary metric with a continuous logit difference before anyone can call the causal null real.

    Add the reserved template holdout your limitations already concede you need.

  3. Great work. An additional experiment that would really help : Run a diff-in-means between two checkpoints trained on unrelated tasks. If that hits AUROC 1, you would have proven the benchmark is degenerate instead of arguing for one

  4. The central contribution is strong. The entry shows a direction that separates each loyalty organism from its matched control at perfect AUROC but is causally inert; the author argues convincingly it detects checkpoint identity rather than loyalty itself, with the valuable lesson that probe accuracy against organisms is gameable. Execution is robust with matched controls, negative categories, one vector used as both probe and intervention. The main limitation, which the author acknowledges, is that the causal metric is saturated. The weakest aspect of the entry is presentation, which is dense. The core story of perfect prediction alongside zero causation is buried under terminology and would land far faster with a plain-language summary and a one-sentence takeaway up front.

Cite this project

@misc{singh2026perfectly,
  title = {{A Perfectly Predictive Loyalty Direction That Does Nothing}},
  author = {Aditya Singh},
  year = {2026},
  month = jul,
  note = {Submitted to Secret Loyalties Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/a-perfectly-predictive-loyalty-direction-that-does-nothing-kldp}},
  url = {https://apartresearch.com/sprints/projects/a-perfectly-predictive-loyalty-direction-that-does-nothing-kldp}
}

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