Skip to content

Report the first failed generated-image evidence check without exposing input data #403

Closed

Description

@mashfromband

Generated-image evidence checks return the same generic failure for an inaccessible evidence file, invalid JSON, an incompatible content review, and failed original provenance. This makes it easy to blame a later condition that was never reached.

Add an opt-in, read-only diagnostic to the existing evidence check. It must use the same guard and verifier, stop at the first failure, emit only fixed stage/reason codes, and never disclose paths, native records, image data or internal errors. It must not register evidence, create keys, relax classification or imply UI approval. Keep the default check/register response compatible.

Acceptance: a healthy synthetic control validates; denied evidence performs zero reads; read/schema/scope/native failures remain distinct; a truthful non-fictional declaration remains rejected by the current contract with an explicit reason; diagnostic output does not register or change original evidence; default CLI behavior is preserved. This diagnoses an unmet qualification and does not claim support for real owner data or resolve any particular asset.

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions