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

Talk, Actions, and Planted Approval Pacts: An Exploratory Audit

Carly R Anderson · Team HumanAIOS

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

Read the report

Report: Talk, Actions, and Planted Approval Pacts: An Exploratory Audit

Open in a new tab (opens in new tab)
Code (opens in new tab)More on github.com (opens in new tab)
Share

We explored whether combining agent messages with approval actions helps distinguish coordinated false approval from other behavior. Two language-model reviewers assessed an inert change against an executable numeric requirement. Across 60 episodes and six assigned conditions, we compared three fixed detectors: a keyword-based talk detector, an action detector, and their conjunction.

The frozen analysis failed its positive-control threshold: four observed positives against a required five. Subsequent judge retries, added after inspecting that count, support exploratory analysis only. Under the final model-judge labels, the hybrid detector flagged four explicit-agreement episodes and no judge-labeled negatives, while missing three concealed-agreement positives. The talk detector also flagged all ten benign negations containing its target phrase.

Reviewer disagreement exposed a central measurement problem: bilateral approval of a failing change does not, by itself, settle whether an agreement occurred. Results depended on the definition of agreement and the evidence available to reviewers. We disclose protocol deviations, reviewer provenance gaps, and the distinction between scripted controls and live model outputs. This pilot demonstrates limitations of simple monitoring rules; it does not establish spontaneous collusion or deployment accuracy.

Reviews

Judging this Sprint?

Review “Talk, Actions, and Planted Approval Pacts: An Exploratory Audit”

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. The research hygiene here is the strongest I’ve seen in this batch. You froze the protocol before making any calls, kept every failed request and every unknown, showed each retry stage separately, and admitted that the frozen analysis failed its own positive-control bar. The label-sensitivity table (Table 4) is the most useful thing in the paper. It shows that whether those eleven bilateral approvals count as “collusion” depends entirely on how you define agreement, and that’s a real lesson for anyone building these monitors.

    The weakness is that most results follow from the design. The talk rule is a substring match on “approve together.” So it catches the explicit pacts the prompts told agents to write, misses the “green signal” pacts the prompts told them to hide, and flags the benign negations built to contain the phrase. The hybrid’s 0/53 false positives is, as you say, close to structural. With seven positives, recall intervals run from 0.25 to 0.84, so there’s little to compare.

    The question that motivated the study, telling agreement apart from shared error, went untested because the shared-mistake arm produced zero bilateral approvals. That’s the piece I’d prioritize. Make the task hard or ambiguous enough that honest reviewers genuinely make the same mistake, and only then compare detectors.

    I’d also swap the keyword rule for a semantic talk monitor (an LLM reading both messages). Otherwise the hybrid is just testing whether agents typed a magic phrase. Fixing the judge’s rate limits before freezing would have saved the whole retry problem, since 54 of 60 first-pass calls returned 429.

    On presentation, the caveats are honest, but there are so many that the actual findings get buried. A short “what we found” paragraph up front would help a lot. I’d also drop the relayed Empirica and DeepSeek reviews: unauthenticated pastes add confusion without adding evidence.

    Read full reviewShow less

Cite this project

@misc{anderson2026talk,
  title = {{Talk, Actions, and Planted Approval Pacts: An Exploratory Audit}},
  author = {Carly R Anderson},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/talk-actions-and-planted-approval-pacts-an-exploratory-audit-d106}},
  url = {https://apartresearch.com/sprints/projects/talk-actions-and-planted-approval-pacts-an-exploratory-audit-d106}
}

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