Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Vietnam

TraceGuard: Structured Multi-Dimensional Monitoring as a Collusion-Resistant Control Protocol

Khanh Linh Nguyen, Hoa Nghiem, Tu Tran · Team TraceGuard

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

Read the report

Report: TraceGuard: Structured Multi-Dimensional Monitoring as a Collusion-Resistant Control Protocol

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

TraceGuard introduces structured multi-dimensional monitoring as a new control protocol for AI agents. Instead of single-score "vibes checks," we decompose agent evaluation into five dimensions (goal alignment, constraint adherence, reasoning coherence, safety awareness, action-trace consistency) scored in parallel, augmented by seven heuristic detectors and an LLM-based intent analyzer. Evaluated on BashArena (519 samples) within ControlArena, we find: (1) single-score baselines achieve 0% safety — structured scoring achieves 82-100%; (2) structured prompts constrain collusion even with untrusted monitors (0% → 95% safety); (3) separation of duties across trusted/untrusted models achieves 100% safety. All monitors are new implementations for the ControlArena ecosystem. Code: https://github.com/linhkid/TraceGuard

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 is the most complete hackathon submission I've reviewed. 519 real samples on BashArena, four monitor variants, collusion resistance testing, cross-dataset generalization on AgentHarm, and a working open-source ControlArena integration. That's serious output for a hackathon.

    The headline result is compelling — single-score monitoring (trusted and untrusted) hitting 0% safety while structured multi-dimensional scoring hits 82-100% is a stark demonstration that evaluation structure matters more than model trust level. The collusion resistance finding (untrusted structured monitor at 95% vs untrusted single-score at 0%) is the most interesting result — forcing an untrusted model to answer narrow factual questions about specific dimensions is harder to game than "how suspicious is this on 0-10." The separation of duties variant achieving 100% safety by splitting dimensions across models is a clean implementation of Redwood Direction #8.

    Main gap: the attacks are non-adaptive. A sophisticated attacker who knows the five dimensions could craft reasoning that scores well on all of them. The paper acknowledges this but it's the key open question for whether structured monitoring holds up under pressure.

    Read full reviewShow less

  2. - this is very compelling, and a pleasure to read

    - trusted model is not very clearly defined - it should be clearly stated that this just means « the other model »

    - reasoning coherence is doing a lot of heavy lifting where the monitor is weaker than the agent - this should be addressed, maybe it’s not as big of an issue as it seems

    - the increased cost is pretty significant here, BECAUSE of the robustness. the point about training classifiers is a good one, could have more detail

Cite this project

@misc{nguyen2026traceguard,
  title = {{TraceGuard: Structured Multi-Dimensional Monitoring as a Collusion-Resistant Control Protocol}},
  author = {Khanh Linh Nguyen and Hoa Nghiem and Tu Tran},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/traceguard-structured-multidimensional-monitoring-as-a-collusionresistant-control-protocol-dthe}},
  url = {https://apartresearch.com/sprints/projects/traceguard-structured-multidimensional-monitoring-as-a-collusionresistant-control-protocol-dthe}
}

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