Calibrate confidence and define the production abstention policy #32

New Issue

timmy · 2026-08-19T01:59:01Z

timmy commented

2026-08-19 01:59:01 +00:00

Parent epic: #4

Outcome

Turn raw scores into a conservative suggestion policy.

Acceptance criteria

Threshold selection is based only on validation data.
Held-out precision among non-abstained suggestions reaches the agreed gate or release is blocked.
Low-quality, unknown, and out-of-distribution images abstain.

Dependencies

#31 — Fine-tune specialist visual classifier with color and quality heads

Required evidence

Test or executable receipt attached to the issue/PR
Privacy/safety boundary checked where applicable
Clean checkout and relevant full suite pass

Delivery

Open a focused branch and PR with Closes #ISSUE. Do not include real medical images in issue comments or Git history.

Parent epic: #4 ## Outcome Turn raw scores into a conservative suggestion policy. ## Acceptance criteria - [ ] Threshold selection is based only on validation data. - [ ] Held-out precision among non-abstained suggestions reaches the agreed gate or release is blocked. - [ ] Low-quality, unknown, and out-of-distribution images abstain. ## Dependencies - #31 — Fine-tune specialist visual classifier with color and quality heads ## Required evidence - Test or executable receipt attached to the issue/PR - Privacy/safety boundary checked where applicable - Clean checkout and relevant full suite pass ## Delivery Open a focused branch and PR with `Closes #ISSUE`. Do not include real medical images in issue comments or Git history.

timmy added this to the M2 — Consented Dataset & Specialist Model milestone 2026-08-19 01:59:01 +00:00

timmy added the

labels 2026-08-19 01:59:01 +00:00

timmy referenced this issue

2026-08-19 01:59:01 +00:00

Export the candidate to ONNX and benchmark CPU, GPU, and device targets #33

timmy referenced this issue

2026-08-19 01:59:03 +00:00