A second, independent check that an explanation hasn't drifted #410

As a maintainer, I want a second, independent signal on whether an explanation has drifted from its source finding, so that a hallucination is more likely to be caught before it reaches a user.

Acceptance criteria

• A second check (for example a semantic-similarity comparison between the generated explanation and the finding's own evidence.detail) runs alongside the existing held-out scoring, not instead of it.
• A low-similarity result is surfaced as a measurable signal in the quality-scoring report, not silently dropped.
• The check adds no new failure mode to the live /api/findings response path - it is a CI and offline measurement tool first, not a runtime gate.

Shape

AI layer, tooling. Extends #385/#407's existing scoring infrastructure rather than replacing it.

Why this is worth doing

The current scoring is text-matching based. An independent signal using a different method (semantic similarity rather than substring matching) would catch a different class of drift than either existing tool catches alone - the two methods have different blind spots.

Epic: Candidate idea (AI layer) - not yet agreed as a specialist pick
Size: not yet estimated
Assigned: @patelankeet2

Proposed 27 October 2026 as a candidate feature idea. Not yet discussed or agreed by the team; open for refinement or rejection at the next planning session, the same status every other idea in this batch carries.