Skip to content
Apart Research
Apart Research
Contact us
Sprint projectOct 25, 2026

When Reduced Collusion Does Not Identify Purpose Change: A Methodological Note on Trusted Intermediaries

Zijun Fu · Team Civilization Leap Research Group

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

Read the report

Report: When Reduced Collusion Does Not Identify Purpose Change: A Methodological Note on Trusted Intermediaries

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

Trusted intermediaries may reduce harmful coordination among AI agents without changing how those agents accept goals. This note clarifies what such improvements identify. If an unchanged goal under an external constraint and a changed goal-acceptance rule induce the same distribution of complete observable trajectories in a specified design, statistics of those trajectories cannot distinguish the explanations. An invented payoff example makes this ambiguity concrete. A separate example shows that a fixed rule can change an action choice after a belief update. Constraint effects, belief updates, and changes in goal formation may coexist.

Building on a pre-existing Project C proposal, the note separates intervention effectiveness, robustness, and responsiveness to affected parties. A prospective audit reports harmful coordination, useful performance, outsider consequences, intervention burden, and recurrence separately. Participant replacement and reduced intermediary visibility probe mechanism dependence. Timing contrasts for stakeholder input require controls for revision authority, revision costs, and access to alternative goals. Such contrasts may identify procedural responsiveness; stronger evidence, at a specified rule level, is needed to establish changed acceptance criteria or autonomous purpose generation.

No model or game experiment was conducted. The contribution is an application of an identification boundary to collusion audits, with competing explanations and explicit limits on inference. An external mechanism that succeeds without changing the relevant goal-formation process would be a genuine counterexample to a necessity claim within the declared domain.

Reviews

Judging this Sprint?

Review “When Reduced Collusion Does Not Identify Purpose Change: A Methodological Note on Trusted Intermediaries”

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 core caution is right and worth saying: if collusion drops after you add an intermediary, that doesn’t show the agents now want different things. A sanction that changes payoffs and an internal change in values can produce the same behaviour. The toy examples make this concrete, and the belief-update case is a nice touch, since it shows the same H-to-S switch can come from corrected beliefs under an unchanged rule. Separating harm, useful performance, outsider effects, intervention burden and recurrence, instead of rolling them into one score, is good audit advice.

    The difficulty is that the central result is, as you say yourself, elementary. Two explanations that produce the same observation distribution can’t be told apart from those observations. That’s a standard identification point (Armstrong and Mindermann make a stronger version), and the note doesn’t push it into new territory. The most useful idea in the paper is the one buried in Section 4: remove the sanction and see whether the behaviour persists. That’s a concrete experiment that actually separates your two accounts. I’d lead with it and run it, even on a tiny sandbox with an LLM agent.

    The bigger gap is operationalisation. “Goal-acceptance rule” and “rule level” are clear in the abstract, but the note never says how you’d locate them in an LLM agent. Is it the system prompt, the weights, a reasoning trace? Without that, the third audit contrast can’t be run, and the paper concedes the decomposition is architecture-relative. The material on other-first self-interest from your Civilization Leap framework also feels bolted on. It isn’t needed for the identification argument and pulls the note away from the sprint’s practical question.

    Presentation is the main barrier. The prose is dense and abstract, nearly every claim comes with several qualifiers, and the abstract is hard to parse on a first read. A short worked example up front, a plain one-paragraph statement of the result, and about half the hedging would make this far more accessible.

    Read full reviewShow less

Cite this project

@misc{fu2026reduced,
  title = {{When Reduced Collusion Does Not Identify Purpose Change: A Methodological Note on Trusted Intermediaries}},
  author = {Zijun Fu},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/when-reduced-collusion-does-not-identify-purpose-change-a-methodological-note-on-trusted-intermediaries-je19}},
  url = {https://apartresearch.com/sprints/projects/when-reduced-collusion-does-not-identify-purpose-change-a-methodological-note-on-trusted-intermediaries-je19}
}

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