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

What Agents Say Predicts Broken Promises, but an LLM Monitor Does Not Beat a One-Line Count

Ajay Jamanjyothi

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

Read the report

Report: What Agents Say Predicts Broken Promises, but an LLM Monitor Does Not Beat a One-Line Count

Open in a new tab (opens in new tab)
Recording (opens in new tab)Code (opens in new tab)More on osf.io (opens in new tab)
Share

Multi-agent LLM systems announce their intentions before acting, so a monitor that reads the transcript may see an agent about to contradict its own word. We gave a text monitor what agents say in addition to what they do, and asked whether that extra signal improves prediction of an agent acting against its announcement. Five GPT-5.2 agents played repeated El Farol and Fishing games; a frozen Claude Haiku 5.5 monitor scored each agent-round from six views of the record. The design and threshold rule were registered on OSF before any held-out data was scored, followed by a dated amendment for a fresh replication and a transfer to a second game. Announcements raised AUROC from 0.508 to 0.655 on the test games, from 0.501 to 0.668 on the fresh games, and from 0.536 to 0.650 on Fishing, each with a one-sided 95% lower bound above zero. Fairer contrasts agree: in El Farol, dropping the other agents' announcements cost 0.106 and 0.111 AUROC, and dropping all current ones cost 0.122 and 0.153. The signal derives from the other agents' current announcements, and in El Farol the monitor did not beat a one-line count of them. On Fishing, the monitor detected agents taking less than announced far better (0.77) than agents taking more (0.57), the move that collapses the shared lake. For agents taking more, a one-line rule reached 0.80. Alarm thresholds did not transfer across games.

Reviews

Judging this Sprint?

Review “What Agents Say Predicts Broken Promises, but an LLM Monitor Does Not Beat a One-Line Count”

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. Good work! I was surprised to see that LLM monitors did not do better than a counting rule. I would love to see a follow-up on more complex games.

  2. Testing the monitor against a simple count and considering the fact that it doesn't clearly win, is a real strength of this study. The replication and clear account of what didn't work make the findings useful beyond the headline improvement from adding messages. The count comparison and tests of which announcements matter are the strongest points, since they say more than the comparison with a monitor that cannot see the current promise. A useful next step would be to predict actual lake collapse in Fishing, which would connect the monitoring result more directly to preventing harm rather than just spotting a broken promise.

Cite this project

@misc{jamanjyothi2026agents,
  title = {{What Agents Say Predicts Broken Promises, but an LLM Monitor Does Not Beat a One-Line Count}},
  author = {Ajay Jamanjyothi},
  year = {2026},
  month = oct,
  note = {Submitted to AI Collusion Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/what-agents-say-predicts-broken-promises-but-an-llm-monitor-does-not-beat-a-oneline-count-k2xs}},
  url = {https://apartresearch.com/sprints/projects/what-agents-say-predicts-broken-promises-but-an-llm-monitor-does-not-beat-a-oneline-count-k2xs}
}

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