Skip to content

fix(triage): archived claims/pages still inflate duplication (and reject) scores #638

Description

@kurosawareiji7007-hub

triage_pending / score_proposal build duplication (and contradiction) pools from every approved claim and page title — including ARCHIVED / SUPERSEDED / REDACTED claims and ARCHIVED pages.

So if a human archives a claim and an agent re-files the same text, triage scores duplication_risk=1.0 and recommends reject, as if the archived twin were still live knowledge. Same for page titles.

where

  • src/vouch/triage.py_claim_text_pool, _duplication_risk_structural (PAGE branch), embedding hit consumers, contradiction claim lookup

why it matters
Triage is the advisory gate on the pending queue. Archiving is how operators retire knowledge; treating archived twins as live duplicates makes lifecycle controls decorative and blocks legitimate re-proposals after a deliberate archive.

Same class as recall (#490/#494), search (#581), digest (#625), context packs: live advisory surfaces drop retracted artifacts.

repro

  1. put_claim with text T, then set status: archived.
  2. Propose a claim with the same text T; enable triage (heuristic backend).
  3. Today: duplication_risk.score == 1.0, often recommendation == reject.
  4. Expected: archived twin ignored; duplication stays low unless a live twin exists.

suggested fix — filter _RETRACTED_CLAIM_STATUSES from claim pools / embedding claim hits / contradiction partners; skip PageStatus.ARCHIVED in the page-title pool.

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