Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 12, 2026Philadelphia

FV-Deception

Lee Wall · Team Leeeeee

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

Read the report

Report: FV-Deception

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

I ran a set of experiments to see if frontier LLM's could intentionally generate formal verification specs that slightly diverge from the natural-language semantic intent, and whether other models could reliably catch this divergence.

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. I am not familiar with the literate on FV, but from what I (+Claude) can tell after digging around there is very little existing work done on this specific problem, and it seems like very important work for the FV field to me. For me this was in shot for a top score on impact if the conclusions were more explicit - what is the so what? What does this really mean and what do we now need to do? It seems like you are very thoughtful about this space, and I was eager to see more of your thinking and takes!

    The execution had room for improvement. The sample size is quite small. The methodology mentions 9 combinations of model attacker/defenders, but only the data and conclusions on using Opus as an attacker was discussed. No analysis of successful / unsuccessful attempts at attacks. I found it hard to really have faith in the (important!) conclusions with these limitations.

    The things that the report communicates are all clean. You know how to write clearly. Well written and argued. I would have liked to see more detail and coverage of the data and experiments that were done.

    I hope you continue with this area of work.

    Read full reviewShow less

  2. Targeting deception in formal specifications is a valuable and neglected direction. The finding that Opus catches fewer of its own bugs than Sonnet is interesting, and I'd like to see whether this replicates with a larger attacker-defender capability matrix. Including concrete examples of deceptive specifications would also help readers interpret what's actually happening.

Cite this project

@misc{wall2026fvdeception,
  title = {{FV-Deception}},
  author = {Lee Wall},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/fvdeception-z2li}},
  url = {https://apartresearch.com/sprints/projects/fvdeception-z2li}
}

More from AI Manipulation Hackathon

All 70 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026