How I keep it honest

Feedback loop

One click never silently rewires production. Toast Δw is ephemeral. I only promote champion weights on GATE PASS — bad labels cannot poison the live model.

ingest / screen → predict (live champion) → Inbox / Quarantine / Call log → human grades: Safe | Block | Needs review → Postgres label (0 / 1 / NULL) → durable feedback buffer → ephemeral partial_fit (toast Δw — this process only) → nightly trainer + held-out golden eval → GATE PASS → promote champion → GATE FAIL → keep old champion

Ephemeral nudge

One SGD step via partial_fit. Proves the wire. Restart discards it. Not the production champion.

Durable champion

Nightly full refit + golden gate (accuracy / FPR / recall). Anti-poison: bad labels cannot promote.

Known-good

Domain match is never enough. SPF/DKIM/DMARC must pass. Spoofed brand From still hits ML.

Report vs grade

Report this is a personal flag + fleet signal. Block / Needs review / Safe still own the label.