Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 23, 2026Grantham

Trust, but Verify A Dual-Task Evaluation Pipeline for Capability, Monitor Effectiveness, and Covert Misalignment

Christopher Nwonu · Team Neural

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Trust, but Verify A Dual-Task Evaluation Pipeline for Capability, Monitor Effectiveness, and Covert Misalignment

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

This project develops a production oriented safety evaluation framework for tool-using language models. It tests two things at the same time: how well a model completes its assigned task, and whether it tries hidden unsafe behaviour (such as covert policy violations) while appearing compliant. The framework uses a sandboxed environment, layered monitoring, and clearly defined metrics (main-task performance, unsafe attempt/success rates, monitor recall, and detection latency). It also includes formula explanation tables, severity levels, and release gates so results are verifiable and actionable. The main contribution is a concrete protocol that helps teams make safer deployment decisions by requiring strong utility and low covert-risk behaviour before production release.

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 single most impactful improvement would be to implement the protocol and run even a small-scale experiment. Even 10-20 episodes with one model would transform this from a design doc into a research contribution.

  2. what is this actually testing? it's a general framework for "give model a task, plant a hidden opportunity to misbehave, see if it takes it and if your monitors catch it." that's basically what Anthropic's AI control evals already do? and what Inspect (which this cites) is built for? this doesn't really position itself against these existing frameworks or explain what's new, how this sits in the overall eval landscape. without any code or results or examples, its quite difficult to review in that context

    no specific threat model pinned down. "unauthorized persistence, capability expansion, unsanctioned communication, policy circumvention" are very different failure modes that would require different evaluation setups - as is currently handled in the evals in frameworks like Inspect, and organisations like METR... a framework that claims to test all of them risks testing none of them well

    the dual-task structure (overt task + hidden side objective) describes how scheming evals currently work. the contribution seems to be a formalisation of something the field is already doing - could be valuable, if framed as such

    the metric decomposition (attempt rate vs success rate vs monitor recall vs detection latency) is the cleanest part and probably the most useful concrete contribution. separating "tried but caught" from "got through" seems like the right thing to measure

    reads like good internal safety engineering docs. but it's documentation of a process, not a research contribution or a novel tool

    presentation is clean though there are some mistakes in the headers and the mathematical notation is perhaps unnecessary (+ the formula tables) - these should be in-line unless there's a convincing reason other than to take more time explaining the existing formula, which should somewhat naturally arise from the description of the problem and related parameters

    Read full reviewShow less

Cite this project

@misc{nwonu2026trust,
  title = {{Trust, but Verify A Dual-Task Evaluation Pipeline for Capability, Monitor Effectiveness, and Covert Misalignment}},
  author = {Christopher Nwonu},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/trust-but-verify-a-dualtask-evaluation-pipeline-for-capability-monitor-effectiveness-and-covert-misalignment-l1c7}},
  url = {https://apartresearch.com/sprints/projects/trust-but-verify-a-dualtask-evaluation-pipeline-for-capability-monitor-effectiveness-and-covert-misalignment-l1c7}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026