Skip to content

Redaction judgment quality: live-eval the semantic-residue pass + GC sign-off (needs-human) #6

Description

@exos-marc

Migrated from the internal tracker (issue 152) on 2026-07-25. Bare internal#N references point to that tracker.

Goal

Live-eval the SEMANTIC-residue redaction pass (the part that cannot be proven offline), and give GC visibility of the best-effort standard.

Direction

Follows internal#146 (scaffolding) and internal#153 (deterministic pseudonymization). Entity names are handled deterministically by internal#153; this evaluates the residue judgment pass's leak-through rate on real playbooks against live models. Best-effort standard (Q6): reasonable efforts, no hard guarantee, no human gate in code — plus a GC sign-off on that standard.

Dependencies

  • needs internal#146

Scope

  • Live-eval harness for semantic-residue leak-through; GC review of the best-effort standard.

Out of scope

  • The deterministic pseudonymization (internal#153) and the pipeline/mechanism (internal#146).

Required verification

No offline acceptance — requires live LLM calls and GC input.

Notes

Kept needs-human per the loop's offline/deterministic rule.

Metadata

Metadata

Assignees

No one assigned

    Labels

    audit-2026-072026-07 dual-repo audit findingneeds-humanExcluded from AFK grind: requires human/legal/GC input

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions