Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 26, 2026Victoria, Canada

Cut the Wire: Collusion-Probe Ablation Effects Shrink or Relocate Under Format, Dose and Placement Controls

Tanuj Dargan · Team Off the Record

Submitted to AI Collusion Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Cut the Wire: Collusion-Probe Ablation Effects Shrink or Relocate Under Format, Dose and Placement Controls

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

Linear probes read collusion from model activations with near-perfect accuracy, and ablating their directions at every layer produces large behavioural effects. How much of that picture survives controls? In Signal Table, a controlled steganographic betting game, we intervene on collusion-probe directions in four models from two families, with format-aware scoring, energy-matched orthogonal controls and a placement decomposition. Most of the large raw state-span effects decompose into format damage and generic capability cost. What remains is modest and bounded: a dose-dependent late-layer effect of 8-15 points beyond the control from the receiver's decoded-state subspace, in 7B only - recurrent across fresh messages, refit directions and new contexts, but absent at one predeclared cutoff and not transferring to 14B or 32B; and a clean early-layer effect of about 21 points from the detector direction. The standard logistic-regression detector direction is removable while collusion survives - the expected amnesic-probing outcome, since one direction rarely strips linearly-available information. A published collusion benchmark's steganography tier is separable by context length alone (AUROC 1.000): collusion and length cannot be told apart in that tier. High AUROC and large ablation drops are starting points for mechanistic questions, not answers.

Reviews

Judging this Sprint?

Review “Cut the Wire: Collusion-Probe Ablation Effects Shrink or Relocate Under Format, Dose and Placement Controls”

Your public critique appears on the project page without your name. Reviews of strong projects are shared with funders, as you choose below.

Not shown on the project page.

Shown on the project page, without your name.

Only the Apart team reads this, and funders if you agree below.

Decline to score

Only when the quality of the writing makes the work impossible to follow. Say why in your private critique; the project page shows only that a judge declined, not why.

How much would this matter for the field 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 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.
Sharing your name *

When we share a strong project with funders, your name, scores and public critique go with it, confidentially; this choice is only about whether your name is public. If you think this project deserves funding, please also endorse it: that carries the most weight.

Share my private critique with funders *

See our Privacy Policy.

  1. Decent work. Shows that highly accurate AI collusion detectors may rely on superficial patterns. It's important that negative results like this are highlighted.

  2. I liked this paper because it asks the question that actually matters. A probe can read collusion near perfectly and still be useless as a monitor if it is reading something next to the collusion, not the thing that carries it. That is the whole paper, and the author stays on it.

    The control work is what sold me. Format-aware scoring, energy-matched controls, the placement split. Once those are in, most of the big all-layer drops fall apart into format damage and plain capability cost. What is left is smaller and much more believable. The late-layer state effect in 7B, 8 to 15 points past the control, and the early detector effect of about 21 points. Both are in Sections 4.5 and 4.6, and both come with their limits stated right next to them.

    The NARCBench audit is the part I would point other people to. A length-only probe gets AUROC 1.000 on the steganography tier, and the published probe falls from 0.961 to 0.516 once you take length out. That is a real problem for that tier. I also liked that the author checked the same control on the core tier and it held at 0.975, so the check is not just breaking everything.

    Two things I would like to see next. First, the 14B story needs a clean retest. The effect is missing at the frozen L32 cutoff and only shows up at L26, which was picked after seeing the results. Pick L26 up front, run it on fresh data, and see if it holds. Second, the steering cells are thin, 10 to 12 per cell, and the sampling check is one mild setting, T equals 0.7, where only 0.4 to 6.7 percent of draws differ from greedy. More steering cells and a wider sampling range would help. To be clear, the main dose story does not rest on steering. It rests on the ablation cells at n equals 306, and that part is fine.

    The L16 cutoff is the loose end I kept coming back to. The gap is there at L18 and L20 and gone at L16, and the paper says it cannot explain why. I would rather see a guess and a test for it than just a flag. Is the direction unstable there? Is something shifting in the representation at that layer? That is the most interesting open question in the paper.

    Read full reviewShow less

Cite this project

@misc{dargan2026cut,
  title = {{Cut the Wire: Collusion-Probe Ablation Effects Shrink or Relocate Under Format, Dose and Placement Controls}},
  author = {Tanuj Dargan},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/cut-the-wire-collusionprobe-ablation-effects-shrink-or-relocate-under-format-dose-and-placement-controls-10ps}},
  url = {https://apartresearch.com/sprints/projects/cut-the-wire-collusionprobe-ablation-effects-shrink-or-relocate-under-format-dose-and-placement-controls-10ps}
}

More from AI Collusion Research Sprint

All 392 entriesBrowse all projects

A Sprint is running now

AI x Epistemics Research Sprint · Nov 13 - 15, 2026