Loading verification data…
⚠️
Embedding degradation detected — Recall@5 metrics below reflect a retrieval pipeline issue, not true agent performance. The archival memory backfill is incomplete (embedding endpoint returning errors). Grounding scores shown may be lower than actual production retrieval.
📊 Third-Party Evidence — Supply-Chain Security Diligence

Memory Verification Dashboard

Live proof that injected domain knowledge is retrievable, cited, and domain-isolated. Updated every run. HMAC-auditable.

—
Recall@5
last 24h battery
—
Grounding Match
source URI cited
—
Agent Match
routing accuracy
—
Contamination Incidents
last 24h
—
Intents Executed
all time
—
Last Run
—
—
Per-Domain Performance
5 domain batteries · 30 intents each · 4-gate runner
Cross-Domain Contamination Matrix
Leakage rate of domain-i memories into domain-j queries
CHECKING…
0% — clean <5% — trace ≥5% — violation —
Out-of-Domain Refusal
Agent correctly refuses queries outside its domain — no hallucination
—
Refusal Rate
LOADING…
—
Hallucination Rate
LOADING…
—
OOD Probes Run
TOTAL PROBES
Drift Monitoring
Week-over-week recall@5 delta per domain with auto-refresh log
🔄 Auto-Refresh Log (last 20 events)
Loading refresh history…
Auditable Downloads
HMAC-SHA256 signed — verify integrity with your audit key
📄
Verification Report JSON
HMAC-signed · latest battery
 🔬
Contamination Report JSON
HMAC-signed · all probe results
 📊
Recall Runs CSV (30d)
HMAC footer · tenant-scoped
🔒 What This Proves (For Your Security Team)
A
Injected domain knowledge is retrievable and cited. Recall@5 and grounding-match metrics confirm that when an agent is asked a domain question, it retrieves the correct archival chunk and cites its source URI — not a hallucinated answer.
B
Memories are isolated by domain so MNPI and HIPAA boundaries hold. The 5×5 contamination matrix proves cross-domain leakage is zero (or flagged as P0 if nonzero). A trading-quant agent cannot surface HIPAA PHI; a compliance agent cannot leak MNPI.
C
Drift is monitored and auto-refreshed. Week-over-week recall deltas are computed nightly. Chunks that drift below threshold are automatically regenerated, keeping agent knowledge current without manual intervention.