Governance[02/04]
Trace / Replay
Memory explainability — replay the full pipeline. A write replays raw → extraction (exact prompt + raw output, kept facts & dropped-with-reason) → fact; a read replays query → candidates → ranking → injected/dropped, with the visible miss-signal.
🏰 MOAT
WRITE
- raw
- extraction
- fact
READ
- query
- candidates
- injection
- kept / injected
- dropped + reason
- no miss
EXTRACT LATENCY412ms
FACTS KEPT / DROPPED1 / 2
RETRIEVE LATENCY38ms
INJECTED / CANDIDATES1 / 3
MISS-SIGNALNO
Capture & sampling
Traces are sampled out in high-volume production to bound overhead. Full-capture is on for debug tenants and is auto-triggered on a miss-signal, so a regression is always replayable. // TODO live-wiring: tenant sampling rate + enable-full-capture toggle