Corroboration

Independent sources asserting the same thing, and the raised confidence that should follow.

Why it matters in a memory layer

The appealing shortcut is to treat textual similarity as evidence of agreement, and it does not survive measurement: on this corpus a refinement scores 0.669, a genuine corroboration 0.505, and a flat contradiction 0.439. No threshold separates them, so a similarity-driven confidence boost grows most certain about the facts it should be doubting. Corroboration is only safe once something has named the relationship — which makes it downstream of conflict detection, not a substitute for it.

Connections

Taught in: From Episode to Belief

Used in: Contradiction Detection · Salience Scoring

Do not confuse with: Deduplication