Eval & Safety
Live Telemetry
Runtime health of the Core Case Router and its four specialists — compute, token spend, inference latency percentiles, and the exact model version serving each agent.
Streaming · 2s · uptime 51h 22m
Runtime signals
CPU utilisation
46.0%
GPU pool 38.0% · 12 vCPU
Memory footprint
64.0%
30.7 GB of 48 GB resident
Token consumption
18.4k
tokens / min · $41.72 spent this hour
Inference latency p50
720 ms
p95 720 ms · budget 2,500 ms
Per-agent runtime
| Agent | CPU | Memory | Latency p50 / p95 | Tokens in / out | Req/s | Errors |
|---|---|---|---|---|---|---|
| Core Case Router | 46% | 62% | 640 ms / 1480 ms | 5.1k / 840 | 8.4 | 0.30% |
| Fraud Agent | 58% | 71% | 910 ms / 2240 ms | 3.5k / 610 | 5.1 | 0.60% |
| Account Agent | 31% | 44% | 380 ms / 820 ms | 1.9k / 310 | 11.2 | 0.20% |
| Knowledge Agent | 63% | 78% | 1120 ms / 2680 ms | 6.2k / 720 | 3.7 | 1.10% |
| Compliance Agent | 27% | 39% | 520 ms / 1180 ms | 2.3k / 390 | 4.4 | 0.10% |
Model version tracking
| Agent | Model | Version | Serving tier | Context | Promoted |
|---|---|---|---|---|---|
| Core Case Router | finassist-orchestrator | v4.2.1active | Gateway · frontier-tier | 200k | 12 Sep 2026 |
| Fraud Agent | finassist-fraud | v3.8.0active | Gateway · reasoning-tier | 128k | 04 Sep 2026 |
| Account Agent | finassist-account | v3.5.2active | Gateway · fast-tier | 128k | 28 Aug 2026 |
| Knowledge Agent | finassist-knowledge | v2.9.4-rc2canary | Gateway · reasoning-tier | 200k | 19 Sep 2026 |
| Compliance Agent | finassist-compliance | v3.1.0active | Gateway · reasoning-tier | 128k | 22 Aug 2026 |
| Knowledge Agent | finassist-knowledge | v2.8.7deprecated | Gateway · reasoning-tier | 128k | 02 Aug 2026 |
Telemetry is sampled every 2 seconds across the orchestration pool. Latency budgets and error-rate ceilings feed the same guardrails used by the failure-mode suite — a sustained p95 above 2,500 ms trips the circuit breaker and routes work to cached fallbacks.