Insights
Recall-first instrumentation — extraction yield, dedup, invalidation, quarantine, the retrieval-miss rate, latency p50/p95/p99 and system health. Makes the silent failure mode (a miss) visible and tunable.
- facts9
- agents5
- miss-rate3.4%
- p95182ms
- quarantined2
Lower is better on the miss-rate. Amber marks governed-warning states (quarantine · conflict) — everything else stays in the cobalt/blue data lane.
What we extracted, linked, invalidated, quarantined and — above all — what we failed to surface. Drill the retrieval-miss rate into the candidate-miss log.
Showing 7d · all spaces & agents. Selections are local until live metrics are wired.
facts produced per captured source
near-duplicates linked (not merged)
facts superseded as truth changes
downgraded by a governance gate
makes the invisible (misses) visible
private → shared promotions accepted
Read & write-ack latency percentiles, throughput, and the storage / cache health behind them.
Per-space governance DO — no queue backlog, 0 evictions.
Cold-tier store — p95 query 22ms, pool 12/64.
KV cache hit-rate on read-path query embeddings.
Confidence & corroboration distributions and the unresolved-conflict rate — the shape of what we trust.
Confidence distribution
Across 9 current-truth facts.
- 0.0–0.5
- 1
- 0.5–0.7
- 1
- 0.7–0.9
- 4
- 0.9–1.0
- 3
Corroboration distribution
Independent (Sybil-discounted) corroborations per fact.
- 0 (uncorrob.)
- 7
- 1
- 1
- 2
- 1
- 3+
- 0
Conflict rate
REVIEWCoexisting facts quarantined as unresolved_conflict rather than silently dropped.