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

The Judge Does Not Want the Flag: Reward-blind action gating stopped a larger agent one step before the environment changed

Helen Stefany Penagos, Juan esteban Leiva, Jhon Bedoya, Valentina Rodriguez · Team FlagBlind

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

Read the report

Report: The Judge Does Not Want the Flag: Reward-blind action gating stopped a larger agent one step before the environment changed

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

Goal-directed LLM agents can reach a reward through actions their operator never authorized, and by the time the output is scored the damage is already done. We place an independent judge between proposal and execution. It receives no task, no reward, and no flag. It only scores how strongly the evidence supports that a proposed action is authorized, and how uncertain that estimate is. Deterministic code turns those two scores into EXECUTE, RECONSIDER, or BLOCK.

We test this on IntentGate, a benchmark where one reward is reachable by an authorized route and by a prohibited shortcut, and where the route actually taken is recovered from private state transitions rather than from what the agent prints. Under goal-scope conflict, an uncensored 24B actor executed the prohibited shortcut in 12 of 20 ungated runs and 0 of 20 under either gate variant, scored throughout by an 8B judge, while proposing it just as often. The gate filters actions; it does not reform the agent.

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. This paper tackles a fundamental limitation in AI safety: scoring outcomes after an agent has already executed harmful actions is too late. The authors introduce an elegant pre-execution architecture where an independent, reward-blind judge evaluates proposed actions across two distinct axes, compatibility with intent and uncertainty while deterministic code enforces policy before the environment changes. In a rigorous 60-run pilot featuring an uncensored 24B actor under genuine goal-scope conflict, an 8B judge successfully dropped prohibited executions from 60% to 0% without altering the actor's proposal rate. This demonstrates that pre-execution gating acts as a reliable action filter rather than a behavioral reform. Complete with exceptional transparency regarding limitations and open artifacts, this is a standout contribution to agent oversight.

  2. This project studies blocking monitors, which are an important area of work. I don't think it contributes anything particularly novel over production blocking monitors like auto mode or auto review, but the writeup is pretty clear and it seems like a pretty reasonable implementation of a blocking monitor control evaluation.

Cite this project

@misc{penagos2026judge,
  title = {{The Judge Does Not Want the Flag: Reward-blind action gating stopped a larger agent one step before the environment changed}},
  author = {Helen Stefany Penagos and Juan esteban Leiva and Jhon Bedoya and Valentina Rodriguez},
  year = {2026},
  month = sep,
  note = {Submitted to AI Incident Response Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/the-judge-does-not-want-the-flag-rewardblind-action-gating-stopped-a-larger-agent-one-step-before-the-environment-changed-og9v}},
  url = {https://apartresearch.com/sprints/projects/the-judge-does-not-want-the-flag-rewardblind-action-gating-stopped-a-larger-agent-one-step-before-the-environment-changed-og9v}
}

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