Fide AI

FID-106 · Open question

Can Independent Reviewers Reproduce a Specific AI Safety Claim?

What is the smallest public evidence package that lets someone outside the original research team reproduce a bounded result, identify where the claim goes beyond that result, or document why reproduction is blocked?

Why the question remains open

A repository can be runnable while an important result remains hard to check. An aggregate may be reproducible from saved judgments even though those judgments have no independent validation. Model access and dependency changes can prevent rerunning the original collection. A claim-level handoff can make the boundary visible and give a contributor a tractable starting point.

Working hypothesis

A proposition to test, not a finding.

An evidence packet that maps one claim to versioned inputs, exact procedures, expected outputs, and known access limits will improve verification under a fixed time budget compared with the released documentation alone. It may add little when the existing artifact already meets a strong reproducibility standard.

Proposed method

How the question could be tested

  • 01Freeze a specific claim, artifact revision, data scope, acceptable numeric tolerance, and time budget before opening the task. Start with one offline aggregate from Fide's released DSEWiki investigation and one independently authored, rights-cleared comparison artifact. Include a blocked or stale dependency case when it can be tested without private access.
  • 02Compare the release documentation with that documentation plus a short claim handoff. The handoff specifies the versioned inputs, entry point, output location, known exclusions and expected result or check. Use separate reviewers or counterbalanced matched tasks to limit learning across conditions.
  • 03Record whether a reviewer can recompute the analysis, rerun the original collection, or repeat the test on independently assembled cases. Track time, author assistance, material discrepancies, and whether the reason for an inconclusive or blocked check can itself be reproduced.
  • 04Let a separate reviewer verify the disposition. Preserve corrections and document whether the problem is missing evidence, execution, arithmetic, adjudication, or a conclusion the available artifact cannot support.

Needed controls

What must constrain the study

  • 01Benchmark against established artifact-review criteria. Treat local Fide materials as self-audit cases and distinguish them from independent review.
  • 02Use pinned, rights-cleared inputs, isolated execution, and bounded resources. Record version drift, missing access and exclusions. Preserve a supported finding of non-reproduction, with uncertainty about the cause where appropriate.
  • 03Separate recomputing a statistic from validating its labels, recreating model outputs, and testing generalization on new evidence.

Relationship to existing work

FID-078 studies validity and transfer of an evaluation. FID-077 studies the evidence needed to reconstruct an incident. This call studies the practical handoff of a published result to an outside verifier. FID-107 asks how to present such a verification task to contributors.

Expected outputs

Artifacts the work should produce

  • 01A claim handoff and one independent reproduction record with exact scope.
  • 02A comparison of completion, cost, and blockers under the two documentation conditions.
  • 03Reusable examples of how to communicate reproduced, discrepant, and blocked results.

Open questions

Uncertainties the protocol must resolve

  • 01Which missing fields cause enough delay or ambiguity to justify maintaining them?
  • 02Which verification steps remain useful when a model or dataset is restricted?
  • 03How can authors distinguish corrections to the documentation from changes to the result?

Open question

Open work

Primary need: reproducibility methods, evaluation engineering, artifact stewardship, and independent review

  • Spend 30–60 minutes recomputing one published DSEWiki aggregate from the pinned public release. Return the revision, Python version, command, counts or exact blocker, and which judgments were held fixed.
  • Review an underlying judgment or propose a comparable artifact from another team.

New to the Contributor Network? Share a short private introduction once, even if you join several calls. You can join the public discussion while introducing yourself.

Join the Contributor Network ↗