classify_zero_shot returns a 0.0-1.0 probability, but the admin UI's
confidence_threshold field took a raw number with no conversion and
no bound -- typing the intuitive "80" (meaning 80%) stored literal
80.0. Since no real confidence score can exceed 1.0, that silently
made every single classification read as below-threshold and fail,
with the failure only surfacing on the next service restart (config
loads once at startup).
Caught live 2026-09-06 before it took effect: the running process
still had the previous config loaded, but the corrupted value was
already sitting in config.local.yaml waiting for the next restart.
Two fixes:
- classifier.encoder.confidence_threshold now validates to [0.0, 1.0]
at config load, matching every other new knob's fail-closed pattern
in this project -- the backstop for any caller, not just the UI.
- The admin field now takes 0-100 (human-natural) and converts to
0.0-1.0 at the save boundary; loading a stored value converts back
to a percent for display.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
