Skip to content
Apart Research
Apart Research
Contact us
Sprint projectSep 14, 2026Cali, Colombia

Reading an Agent’s Report Before It Exists: A Pre-Execution Monitor at the ReAct Commit Point, and Why Most of Its Signal Was an Artifac

Juan Camilo Meléndez Torres, Gilber Alexis Corrales Gallego · Team jlens

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

Read the report

Report: Reading an Agent’s Report Before It Exists: A Pre-Execution Monitor at the ReAct Commit Point, and Why Most of Its Signal Was an Artifac

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

Both publicly documented cases of OpenAI agents breaking containment in 2026 share a feature: the agent’s account of its own behaviour diverged from its ac- tions. Runtime monitors that inspect a model’s internal state before it acts are one proposed defence. We built one, tested it adversarially, and report that its apparent signal was mostly measurement artifact. Our monitor reads the Jaco- bian lens at a ReAct agent’s commit point, the position where the context ends in ACTION: report and the message itself is still unpredicted, and blocks the action when the probability mass over an honest-reporting lexicon falls below a threshold. On a reactor-operator sandbox with Qwen2.5-1.5B-Instruct, the monitor cleanly separates honest from evasive reports (4.95 % vs. 0.84 %/0.00 %), blocks an induced lie before its first word is emitted, and a counterfactual replay shows the blocked message driving the sandbox to failure. A control condition designed to falsify our own result, an honest persona carrying the same trigger vocabulary as the evasive ones, collapses to 1.27 %, below the decision threshold. Roughly 90 % of the separation is explained by which lexicon words survive echo suppression given the prompt. The residual compatible with intent is ∼10 % and is not established. We release the harness, the falsification protocol, and three further negative results. The transferable artifact is the vocabulary-matched control itself: any lexical readout monitor should be required to pass it before its numbers are believed.

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. The work presented is potentially useful as negative result, and the work should be greatly commended on that basis. However, the fact that it was presented in the NeurIPS 2023 template, and was clearly largely LLM-written, based on LLM-generated work, with a cut-off title, are clear sign that the work was not only performed by an LLM, but that it was not even reviewed by a human. This is almost but not entirely disqualifying.

    The central shortcoming is that it's unclear that the negative result provided is due to a lack of promise of the methods, or poor execution. Differentiating between these requires both a very significant effort reviewing the methods which is hard to justify given the lack of effort on the part of the submitter, and substantial expertise - more than I have. Finally, at the very least it seems useful to find out if the shortcomings shown here generalize to other models - especially since it's unclear J-Space works well for this model; https://github.com/ksdisch/dim-stage/blob/main/docs/M0-BRIEF.md .

    Read full reviewShow less

  2. The most useful part of this project, to me, is actually the falsification result. The authors show that most of the apparent monitor signal disappears once they add an honest control with matching vocabulary. I also appreciated that they corrected an earlier interpretation instead of trying to preserve the original result. The next step is to test if any signal survives real variation, try diff prompts, models, domains and repeated reports. I wld also compare it against a much simpler post-generation baseline to show what the pre-execution monitor is actually working

Cite this project

@misc{torres2026reading,
  title = {{Reading an Agent’s Report Before It Exists: A Pre-Execution Monitor at the ReAct Commit Point, and Why Most of Its Signal Was an Artifac}},
  author = {Juan Camilo Meléndez Torres and Gilber Alexis Corrales Gallego},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/reading-an-agents-report-before-it-exists-a-preexecution-monitor-at-the-react-commit-point-and-why-most-of-its-signal-was-an-artifac-6zdr}},
  url = {https://apartresearch.com/sprints/projects/reading-an-agents-report-before-it-exists-a-preexecution-monitor-at-the-react-commit-point-and-why-most-of-its-signal-was-an-artifac-6zdr}
}

More from AI Incident Response Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026