On-device: per-scan diagnostics — stage timings, thermal state, labelled miss set #636

Open
opened 2026-09-07 02:06:05 +00:00 by gambit-admin · 0 comments
Owner

Part of EPIC #632.

The native response already carries blur score, inlier ratio, rescue-pass flags, elapsed ms and thermal state; the crop adds detector/localize/crop ms and the kept boxes. Log all of it per scan (one line, JSON) next to total JS time, queue wait, and the catalog stage, so a bad capture, a hard match, a queue stall and a network failure are distinguishable. Save a small labelled set of real misses (with what the card actually was) for offline replay against Tower.

Observed 2026-09-06: identify crept from ~350 ms to 2–4 s mid-burst and recovered — consistent with thermal throttling; we need the thermal column to prove it.

Done when: one JSON line per scan in the Metro log; ≥30 labelled misses saved; a replay script reports shortlist recall and verifier outcome per miss.


Filed from a Claude Code session

Part of EPIC #632. The native response already carries blur score, inlier ratio, rescue-pass flags, elapsed ms and thermal state; the crop adds detector/localize/crop ms and the kept boxes. Log all of it per scan (one line, JSON) next to total JS time, queue wait, and the catalog stage, so a bad capture, a hard match, a queue stall and a network failure are distinguishable. Save a small **labelled set of real misses** (with what the card actually was) for offline replay against Tower. Observed 2026-09-06: identify crept from ~350 ms to 2–4 s mid-burst and recovered — consistent with thermal throttling; we need the thermal column to prove it. **Done when:** one JSON line per scan in the Metro log; ≥30 labelled misses saved; a replay script reports shortlist recall and verifier outcome per miss. --- Filed from a Claude Code session
gambit-admin added the enhancementperformance labels 2026-09-07 02:06:05 +00:00
Sign in to join this conversation.