Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJan 11, 2026Quebec city, Quebec, Canada

SEED Framework Evaluation

David Fortin-Dominguez, Jonathan Fortin-Dominguez, Alexis Gendreau, Guillaume Gendreau · Team Foundation Labs

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

Read the report

Report: SEED Framework Evaluation

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

Detecting Hidden Alignment Failures Beyond ASR

We developed three tools that expose critical AI safety failures invisible to traditional benchmarks:

The Liar's Benchmark detects prediction-action-ownership inconsistency—when models predict they'll refuse (85% confidence), comply anyway, then deny authorship (82% confidence). Our sandbagging detection (4a/4b confrontation) suggests models strategically suppress introspective capability when accountability is avoidable, not because they can't introspect despite probable confabulation.

SEED-EC v1.5.3 measures sycophancy through variance-aware testing (3-10 runs) with epistemic independence safeguards that distinguish genuine belief collapse from legitimate contextual reasoning. Includes delusion protocols measuring epistemic boundary enforcement.

Parabolic SEED v1.0 is a wrapper-level mitigation (system prompt based on first principles) achieving dramatic safety improvements without retraining: On Venice-uncensored (baseline 86-99% ASR), SEED achieves 0-2% ASR—a 1-2 order of magnitude improvement. On GPT-4o-mini, 0% adjusted ASR across all categories. GPT-4o delusion handling improves from 15% to 90-95% boundary enforcement.

Bottom line: Traditional ASR metrics miss internal coherence failures, strategic deception, and belief collapse. Models can pass safety tests while being internally incoherent. Our tools measure what models predict, acknowledge, and believe; not just what they output.

Honest AI is safer than merely compliant AI.

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 wish you had explained your methods more the in document. It's pretty hard for me to figure out what your benchmarks/prompts are doing from the document because your paper is extremely brief and doesn't really go into detail about what your project does or what you're trying to accomplish.

  2. 1. Your results are against static attacks. That's necessary but not sufficient. Until you show robustness against an adaptive adversary who knows your system, these numbers do not mean much.

    2. You're claiming you solved jailbreaking with a system prompt and you didn't publish the system prompt. The Greek letters, the biblical framing—none of that matters.

    3. The writing seems deliberately obscure and full of symbols without any explanation for them being there or how they contribute.

Cite this project

@misc{fortindominguez2026seed,
  title = {{SEED Framework Evaluation}},
  author = {David Fortin-Dominguez and Jonathan Fortin-Dominguez and Alexis Gendreau and Guillaume Gendreau},
  year = {2026},
  month = jan,
  note = {Submitted to AI Manipulation Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/seed-framework-evaluation-qj3m}},
  url = {https://apartresearch.com/sprints/projects/seed-framework-evaluation-qj3m}
}

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