ADR-0044: Every finding names its level of proof, in every format

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, ADR-0033, ADR-0040, ADR-0041 Trigger: the owner, 23 October: "we have to show something more."

Decision

1. Each finding carries one named level of proof, from strongest to weakest:
   • proved by a test: passes on the old code, fails on the new (ADR-0002);
   • proved by a tool: a deterministic tool's result differs between old and new (ADR-0040);
   • proved by data: a query and the rows it returned (ADR-0002);
   • agreed by two models, not proven (ADR-0041).
2. The level is shown first in text, Markdown, HTML and the pull-request comment, and is a field in JSON and SARIF. A proposed fix carries its own level: fix proved by the same test or fix not proved (ADR-0042).
3. The totals count by level, so a reader sees at once how much is proven and how much is opinion.

Consequences

• One word tells a reader how far to trust a finding, without reading its evidence.