Skip to content

Reality probe: should evidence provenance be a gate before claim semantics? #41

Description

@hippoley

Related external issue:
openai/codex#40755

@xormania — your reproduction is interesting to CounterProof for a reason slightly different from the Codex bug itself.

You already did the hard work: the reported review can cite a stable-looking Git SHA as command evidence even though that object does not resolve in the repository, while the review's own Reviewed commit: pin does resolve.

CounterProof currently carries exact BASE / HEAD SHAs and a digest in its evidence artifacts, but it mostly treats those as provenance metadata around behavioral evidence.

#40755 suggests provenance may need to be a gate, not decoration.

The evidence state I am trying to name

Something like:

claim:
  "commit violates attribution invariant"

evidence citation:
  git cat-file -p <sha>

provenance:
  cited SHA resolves in reviewed repository?  NO
  cited SHA == reviewed commit?               NO
  reviewed commit itself resolves?            YES

claim status:
  INCONCLUSIVE / provenance-invalid

In other words, CounterProof should not even reach “is the claim witnessed / contradicted?” if the object named as evidence cannot be tied to the reviewed history.

Why I am asking instead of implementing

We just learned from external reviewers that evidence axes are useful only if the vocabulary stays mechanical and auditable. I do not want to add another generic “trust score.”

A possible provenance contract would be narrowly deterministic:

  • cited commit/object resolves;
  • object belongs to the repository under review;
  • object identity matches the declared HEAD / BASE / reviewed commit when that relationship is claimed;
  • command evidence names the same object its prose conclusion describes.

Question

From the failure mode you measured in #40755, would a compact provenance-valid / provenance-invalid / unverified gate have made the bad findings immediately actionable for you, or is the more useful artifact simply “show the exact command + stdout + reviewed SHA” and let the reviewer compare them?

Either answer helps decide whether CounterProof should treat provenance as a first-class evidence axis or leave it as raw receipt metadata.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions