[LOW] Decision: does an evidence-support lint belong in the CLI, or upstream at staging? (PR #59) #63

BLUF — what needs deciding

Does an evidence-support lint belong in the CLI at all, or upstream at staging time?

• PR [LOW] feat(review): lint findings whose confidence outruns their evidence #59 adds vhir review --findings --lint, flagging findings whose stated confidence outruns their cited evidence: HIGH confidence with no audit_ids, an interpretation asserting intent or attribution with nothing cited, an interpretation far longer than its observation, HIGH confidence alongside hedging.
• The underlying problem is real and is not in dispute: confidence and confidence_justification are written by the same generative model that wrote the finding, and the examiner is shown a self-report as though it were calibrated.
• The honest limitation is the decision. This is a lint over surface features of the text. It does not and cannot judge whether the cited evidence supports the claim — that needs reading the artifacts. A flagged finding may be perfectly sound; an unflagged one has not been verified. Both caveats are printed with the output, but a reader can still over-trust it.

Options

1. Ship as-is, in review. Cheap, opt-in, honest about its limits. Risk: an examiner reads "nothing flagged" as "checked".
2. Move it upstream to staging. The MCP backend writes these findings; flagging there means the model gets told its confidence is unsupported before a human ever sees the finding. Arguably the right place — the CLI is then reviewing better input rather than annotating bad input.
3. Don't ship it. If you judge that a surface lint invites more false confidence than it removes, that is a defensible call and I would not argue hard against it. This is the weakest of the four proposals in this series and I would rather say so than oversell it.

If you ship it

• Should flags feed [MEDIUM] feat(review): rank the review queue by investigative value #57's ranking, sorting flagged findings up? The disagreement between a writer's stated confidence and an independent check is the informative signal. Not wired; the two PRs are independent.
• The word lists (_INTENT_TERMS, _ATTRIBUTION_TERMS, _HEDGE_TERMS) encode English, Windows-centric IR prose. They are plain tuples at the top of the module.
• attribution-claim fires whether or not evidence is cited, on the grounds that a single host's artifacts essentially never support attribution. Tell me if that is too aggressive.

Related: #56, #57, #58 — same design review, each independently cherry-pickable.