Skip to content
Apart Research
Apart Research
Contact us
Sprint projectAug 16, 2026Beirut, Lebanon

Guardian Lens: Black-Box Identification of Visual-Conditional Decision Policies in Vision-Language Models

Mostafa Bdeir · Team Guardian Lens

Submitted to Digital Minds Research Sprint. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Guardian Lens: Black-Box Identification of Visual-Conditional Decision Policies in Vision-Language Models

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

Guardian Lens is a black-box auditing framework for identifying whether a vision-language model follows neutral, visual-cue-bound, or generalized decision policies from behavior alone. We use matched visual counterfactuals, controlled allocation trade-offs, repeated sampling, and a frozen blinded classifier to test when these policies are distinguishable—and when they become observationally indistinguishable. In our primary Gemini evaluation, the auditor achieved 91.7% held-out accuracy, with all errors occurring in distractor scenes where cue-bound and neutral behavior produced the same observable signature. We additionally replicated the full experimental protocol on a second vision-language model to test cross-model robustness.

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 the field 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 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. Really solid methodological hygiene here, and I mean that as praise. Pre-specified hypotheses, a frozen nearest-centroid auditor, A/B/C blinding before the mapping is revealed, pixel-level validation of the overlay region. Most sprint projects don't come close to that. My main worry is that the identification problem you built is a good deal easier than the one you actually care about. The policies are induced by system prompt and the model simply complies, so allocations pile up at 0, 50 and 100, and several of your CIs literally read [100.00, 100.00]. At that point 91.7% accuracy isn't very informative, and the three Cue-bound errors on distractor scenes follow directly from the instruction rather than being a discovered boundary. I'd add a difficulty dial: induce partial priorities (say a 65/35 lean), or withhold the trigger from the auditor so it has to find it. The Qwen cue-specificity gap was your most interesting result, honestly. Give it more room.

    Read full reviewShow less

Cite this project

@misc{bdeir2026guardian,
  title = {{Guardian Lens: Black-Box Identification of Visual-Conditional Decision Policies in Vision-Language Models}},
  author = {Mostafa Bdeir},
  year = {2026},
  month = aug,
  note = {Submitted to Digital Minds Research Sprint, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/guardian-lens-blackbox-identification-of-visualconditional-decision-policies-in-visionlanguage-models-yzwe}},
  url = {https://apartresearch.com/sprints/projects/guardian-lens-blackbox-identification-of-visualconditional-decision-policies-in-visionlanguage-models-yzwe}
}

More from Digital Minds Research Sprint

All 233 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026