Shared Knowledge

A document covering six things goes stale in one of them and stays live in the other five.

Across three corpora, changing chunk strategy moved answer quality more than changing the embedding model.

Records carrying a source were acted on directly; records without one triggered a fresh investigation.

Sizing charts, pricing tests and editorial formats all showed an early reading that later reversed.

Two abandoned approaches were independently re-proposed within six months of being dropped.

An agent with no record asks. An agent with a stale record acts, and acts wrongly.

Full-corpus access made answers worse, not better — relevant records were crowded out by plausible ones.