ADR-0041: A problem two models agree on, without proof, is shown apart and labelled as an opinion

Status: accepted · Date: 2026-10-23 Decided by: the owner · Written by: the advisor (claude-code, claude-opus-5-5), 2026-10-23 Follows: ADR-0002 (evidence), ADR-0037 (the second reviewer) Trigger: the owner, 23 October: "or some other verification, not from a database, but from judgement, the way we do it with our reviewers?"

Context

Proven reports a problem only with evidence: a test, a query, a tool's run. Some real problems cannot be proved that way in one review, for example a confusing design or a missing check that no test exercises. With two reviewers from different model families (ADR-0037), their independent agreement is a signal worth showing, but it is not proof.

Decision

1. A finding that both reviewers name independently, at the same place, and that no evidence proves, is shown in a separate section: Agreed by two models, not proven.
2. It is never mixed with proven findings: it does not fail a pipeline, does not count in the totals, and carries both models' reasons word for word.
3. One model alone is not enough: a claim only one reviewer makes stays dropped, as before.
4. JSON and SARIF mark these with their own level (note), so a machine never reads them as proven.

Consequences

• The reader gets the judgement of two independent reviewers without Proven claiming more than it can prove.
• Needs a second reviewer; with one, the section is empty and says why.