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

Constraint Dilemma

Manu Xaviour Thaisseril Shaju, Soham Takawadekar · Team Manu- Soham

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

Read the report

Report: Constraint Dilemma

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

Trusted monitors can catch backdoors in AI-generated code, but only by flagging a large share of honest submissions for costly human review. We propose enforced comment constraints — mandatory, structured documentation rules that create a dilemma for adversarial models: follow the constraints, which bind comments to actual code semantics, exposing the backdoor to the monitor — or break them and get caught by an automated compliance checker before the monitor even runs. Building on Sreevatsa et al. (2026), who showed that optional comments improve trusted monitoring, we make comments mandatory and structured, eliminating the attacker's ability to write vague or misleading documentation. We introduce a 5-class constraint taxonomy ranging from free AST-based presence checks to LLM-verified semantic accuracy, and evaluate it against six adversarial commenting strategies across 200 coding problems. Our experimental data is synthetic due to compute limitations; the contribution is the constraint dilemma framework and methodology, not the specific numbers

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 constraint dilemma framing is clever — forcing adversarial models to either honestly describe their backdoor (exposing it to the monitor) or violate the format and get caught by an automated checker before the monitor even runs. No third option. The 5-class taxonomy is well-structured, and the observation that Classes 1-2 (AST parsing, regex) are free and deterministic makes the cheapest version immediately deployable. Even just requiring branch-level comments eliminates the entire "leave the backdoor uncommented" strategy class at zero cost.

    The main issue is that the entire evaluation is synthetic — no real model outputs, no real constraint checking, no real monitoring. The authors are upfront about this (GPU availability), which I respect, but it means the heatmaps and coverage matrices are LLM-generated simulations of what results might look like, not evidence that the dilemma actually works. The backdoor model tested (if n == 47: return -1) is also very simple — real backdoors look like honest mistakes, not obvious triggers, and it's unclear whether the constraint taxonomy holds against adversaries who write comments that are literally true but selectively emphasize the wrong things.

    Good framework, solid writeup, needs real validation on ControlArena to land.

    Read full reviewShow less

  2. Nice framing for enforcing comment standards to support control protocols and monitoring. Due to compute constraints the current results are unfortunately simulated and uninformative, but the approach would be interesting to explore in practice. In particular measuring the increase in FPR from the constraints, and the ability of a powerful model to produce adversarial comments that both pass constraints checks and fool the monitor.

Cite this project

@misc{shaju2026constraint,
  title = {{Constraint Dilemma}},
  author = {Manu Xaviour Thaisseril Shaju and Soham Takawadekar},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/constraint-dilemma-udka}},
  url = {https://apartresearch.com/sprints/projects/constraint-dilemma-udka}
}

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