0.8.0 — 2026-08-08
- Measure recognition from the UI (Settings → "Does it actually work?"). Until now the
calibration analysis lived inscripts/, which needs shell access — so Home Assistant
app users could change the threshold,match_top_kandmin_face_px, but had no way to
see whether it helped. That is the opposite of the advice this project gives everyone
else. Closes #7. - Two measurements, both without manual labelling: leave-one-out over the gallery
(each reference photo tested against the gallery it was removed from — real ground
truth) and a probe over recent events showing how much headroom each recognition
has above the threshold. - The result states the one number that decides whether the threshold can be lowered:
how high a stranger got. Ignore anchors and cross-person matches both count. If any
misassignment shows up, it says so instead — then the threshold goes up, not down. - Events whose face is already in the gallery are excluded from the probe; they score
~1.0 and would flatter a lowtop_k. - The analysis now lives in
app/analysis.py, so UI and CLI compute identical numbers.