unsay bench · 200 revisions · 2026-09-28T07:57:53.121Z
transport: Streamable HTTP over the public internet → https://unsay-production.up.railway.app
client: this machine · sdk @modelcontextprotocol/sdk 1.30.0
path: POST /write (HMAC-SHA256 verified) → store → notification → authorized re-read
at-rest: as deployed (see https://unsay-production.up.railway.app/verify)

segment                             p50        p95        max
signed write → notification     251.9ms    332.8ms    436.0ms
notification → re-read          274.8ms    399.9ms    879.5ms
─────────────────────────────────────────────────────────────
END-TO-END (write → value)      533.3ms    667.9ms   1148.3ms

speech window: a 12-word utterance ≈ 3400 ms (an assumed speech rate, not a
measurement of Alexa+ TTS)
retraction lands mid-sentence in 200/200 runs (100%)
a host was subscribed for every run: yes

NOTE: two real network hops are inside every figure — the clinician's write
travels client → server, and the notification and re-read travel server → client.
Both ends are the same machine here, so the round trip is paid twice; a clinician
and a host in different places each pay one leg. Not a measurement of Alexa+.