20 August 2026 · Week 01
Field test week 01: ten units warm, first numbers
Ten LIZ units are up on our own hardware for the first week of live field testing. Here is what we measured, what broke, and what we fixed.
Fleet Status
We are running a ten-unit LIZ test fleet on our owned hardware. Seats are promo-gated for the pilot — the live count is on the landing page. Every unit is an isolated OS process with its own dedicated SQLite database and Qdrant memory collection.
What Broke
Under sustained multi-turn dictation, client-side retry timers on Android occasionally doubled when a background task woke late, causing unnecessary duplicate turns.
What We Fixed
Hardened the exponential backoff jitter on the Android APK and added cross-platform SHA-256 crash grouping on the Hub to catch regressions within seconds.
One Number We Watched
Speech recognition stays under half a second at p95 on dedicated GPU hardware, with full LLM turns finishing in approximately two seconds. Speech-to-text latency clocked at 332 ms p50 (431 ms p95), while full voice turns completed with a typical reply latency of around 2 seconds.
Next Week
Distributing referral codes to early pilot testers, watching Qdrant memory retrieval query latencies under load, and verifying automated daily owner telemetry digests.