Targets, not an SLA
All numbers below are targets from the internal latency budget, not measured production results. Integrated statistical WASM qualification remains incomplete; preliminary statistical tail results exceeded the target. Network and upstream time are separate from engine overhead.
| Path | Target | Validation boundary |
|---|---|---|
| WASM SDK | < 100 µs added | Target; local default-engine paired-p50 fixtures do not qualify all statistical paths. |
| Sidecar | < 1 ms added p99 | Target; separate Unix HTTP/direct loopback release-default-engine benchmark. |
| Shape diff | p50 < 50 µs; p99 < 500 µs | Target; named shape_diff native fixture excludes parsing/setup. |
| Statistical detectors | p50 < 200 µs; p99 < 2 ms | Target; field_detectors is a bounded native subset. |
| Availability gate | < 10 µs | Target; run-availability.mjs checks mean confidence bounds, not p99. |
Interpret evidence
A unit test does not establish latency. A report must identify runtime, hardware, build, workload and included costs. There is no hosted latency or uptime guarantee. Registry refresh runs asynchronously outside the call path.
Specification
Read the published modes guide and the guard loop.