Latency lab
Measured, not claimed. Fire real queries at the running retrieval layer and see how much of a spoken sentence the shield needs before it can answer. Then read the committed evaluation over eighteen full call transcripts.
Live benchmark
Engine search p50
-
Embed + search p50
-
Embed + search p95
-
Browser round-trip p50
-
Share of one spoken sentence (~2500 ms) consumed before the shield has an answer-
Retrieval measured on this server. Illustrative comparison only: a 350 ms network request would consume 14% of this assumed sentence duration. This does not measure speech-to-warning latency.
Detection evaluation
No committed evaluation found. Run npm run eval.