US-61 — Measured confidence and cited justification for explanations #394

As someone reading an explanation, I want to see which evidence each claim rests on and how often explanations of this kind are right, so I know how far to trust it.

Acceptance criteria

• Justification: each claim in an explanation names the finding field it came from
• Confidence: a measured accuracy per finding type, from A held-out, repeatable score for the explanation layer (#346) #385's held-out set, never a figure the model states about itself
• Both travel outside the F-1 finding, the way explanation_source does, so the finding format is unchanged
• Tested without needing Batfish or Ollama

Shape

Not a pipeline feature: AI layer.

Why this is in the milestone

A model's own stated confidence is poorly calibrated; a number next to a security finding misleads in either direction unless it is measured.

Depends on #385.

Explicitly not in scope

Confidence stated by the model.

Epic: Specialist pick (explainable and trustworthy AI for network security) | Size: estimated at sprint planning | Assigned: @patelankeet2

Specialist pick agreed by all four on 13 October 2026 (slate row 4). Acceptance criteria are drafted from the slate's wording; the owner refines them at sprint planning on 19 October.

Part of Phase 3 - market readiness.