corrective panel transfer

Labnote 010: corrective panel transfer🔗

Question🔗

Does adding an approved Inventory-only and closed-panel development batch fix the cross-batch false-positive behavior of the frozen MobileNet V3 Small panel classifier without spending a fresh holdout?

Method🔗

• Private input: four approved, structurally masked development batches only.
• Corrective batch: 12 Small-UI observations containing eight Inventory-only

states and four closed-panel negatives.

• Spent and future holdouts: not transferred to the execution environment and not read by the

experiment.

• Model: ImageNet-pretrained MobileNet V3 Small with frozen backbone;

classifier head only.

• Augmentation: bounded translation, scale, brightness, contrast, and

saturation.

• Evaluation: four leave-one-batch-out folds, 18 epochs per fold.
• Accelerator: RTX 4070 through a non-preemptive accelerator lease.
• Internal execution record retained; identifier withheld.

Result🔗

Aggregate exact accuracy was 29/42 (69.0%). The job completed on CUDA with exit code 0 and no stderr. The scheduler and compute broker remained healthy with zero restarts, and the accelerator lease was released normally.

Decision🔗

Reject this corrective model and do not spend a fresh holdout. The new batch exposed rather than repaired cross-environment confounding: the model produced four false Hero detections in a batch containing no Hero examples, missed two Inventory states there, and regressed to four false Inventory detections on the first batch. No fold reached 80% exact accuracy.

Do not respond with more epochs or another blind classifier-head run on the same corpus. The next development slice should separate panel evidence into localized regions, add visually diverse negatives for each class, and freeze a development-only calibration gate before any new holdout is collected or read.