Skip to content

Open question: leakage mask-scoring gate needs a quieter kernel #57

Description

@IanMayo

Recorded in specs/108-advisories/spec.md ("Deliberately not in this feature") and task T708: the measurement geometry travels complete in the run-manifest sibling — identification radius, every sampled position, the interval — but the change mask between successive released products is not scored, because the shift-advect kernel's per-cell noise moves every cell between runs. Against that field the mask comparison scores at chance for the wrong reason, and a gate nobody can fail is worse than none (the V1 FR-015/FR-017 lineage).

The geometry's shape needs no change: a V3 backend with a quieter kernel (or a V2 kernel variant with per-cell noise suppressed for the scoring run) can implement the gate against the sibling documents as they already stand. E11's producer-parity test already holds the radius.

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions