Found during the resident-brain-vs-cold-start assay (docs/audits/2026-08-22-resident-brain-assay.html, §02/§06). Kept out of that assay's A/B verdict because it's pre-existing and independent of which cognition backend is compared — this is about the resident brain's own tail latency against its own declared limit.
The numbers
voice_turn declares a 45s budget internally.- In the assay's controlled single-variable test (14 identical prompts against
spark-brain), tail latency reached up to 86.5s — 1.9x the declared budget — with p90 at 39.3s and mean 23.0s (median 18.8s).
- Outside the controlled test: 517 timeout/unavailable events logged since 2026-08-17, 11 of them on the day of this test before testing even began.
Where
src/pxh/brain.py — ask_brain() / the voice_turn call path and its budget accounting.
Open question (from the assay's own falsification section, §07)
Does the elapsed time correlate with:
- concurrent reflection/other-kind traffic contending for the single-flight
FileLock on the session (see ask_brain meter, state/brain/meter.json per kind per day), or
- session context size approaching a recycle boundary?
The assay's synthetic back-to-back calls may themselves induce contention that a real week of family-use traffic wouldn't show at the same rate — so the recommended next step is instrumenting the real base rate and correlating it against those two candidate causes over a week of actual use, not more synthetic load.
Found during the resident-brain-vs-cold-start assay (
docs/audits/2026-08-22-resident-brain-assay.html, §02/§06). Kept out of that assay's A/B verdict because it's pre-existing and independent of which cognition backend is compared — this is about the resident brain's own tail latency against its own declared limit.The numbers
voice_turndeclares a 45s budget internally.spark-brain), tail latency reached up to 86.5s — 1.9x the declared budget — with p90 at 39.3s and mean 23.0s (median 18.8s).Where
src/pxh/brain.py—ask_brain()/ thevoice_turncall path and its budget accounting.Open question (from the assay's own falsification section, §07)
Does the elapsed time correlate with:
FileLockon the session (seeask_brainmeter,state/brain/meter.jsonper kind per day), orThe assay's synthetic back-to-back calls may themselves induce contention that a real week of family-use traffic wouldn't show at the same rate — so the recommended next step is instrumenting the real base rate and correlating it against those two candidate causes over a week of actual use, not more synthetic load.