JARVISLabs

Reflex lane first audio, p50

Readings from different voice paths are not directly comparable; each point names its path.

Reflex lane first audio, p50

Then 82.1 ms 20 Aug 2026 Now 403.5 ms 17 Sept 2026 worse

  • measured or observed

budget p95 300 ms

Show the numbers
Date Value Evidence Revision Samples Note
20 Aug 2026 82.1 ms MEASURED 28fc5cb6 80 Successive readings of the instant-answer lane's time to first sound, each checked by a planted-delay control.
12 Sept 2026 395.2 ms MEASURED 5e8175f2 16 Successive readings of the instant-answer lane's time to first sound, each checked by a planted-delay control.
17 Sept 2026 403.5 ms MEASURED e00c6f1b — A quick acknowledgement starts playing in about 0.4 s (fallback voice).

Readings from different voice paths are not directly comparable; each point names its path.

Test: tools/probes/probe_reflex_first_audio.py: wall clock from return of the injected transcribe callable to the real speaker's first audio write.

Protocol

Test definition

tools/probes/probe_reflex_first_audio.py: wall clock from return of the injected transcribe callable to the real speaker's first audio write.

Limits

Limitations

  • Earlier probe generation on an earlier audio path; the handoff later calls such readings stale Piper-era artifacts; NOT comparable with post-lm/1 readings.
  • post-lm/1 instrument (planted 40 ms control recovered ~95-103%); injected transcript, real speaker; superseded by the next retake; the later 403.5 ms reading is the parent's headline.
  • Measured with the older Piper voice, NOT the owner-approved C013 voice; C013 audible first-audio is still unmeasured.

Log

Recorded values

Date Value Evidence Samples Source
20 August 2026 82.1 ms MEASURED 80 evidence/latency/reflex_first_audio*.json (gitignored, preserved prior artifacts)
12 September 2026 395.2 ms MEASURED 16 evidence/latency/reflex_first_audio*.json (gitignored, preserved prior artifacts)
17 September 2026 403.5 ms MEASURED — docs/orders/post-lm/HANDOFF.md (2026-09-17)

Source paths refer to the private JARVIS repository and runtime records. They are listed for audit and are not published.