localized panel transfer

Labnote 011: localized panel transfer🔗

Question🔗

Can class-specific localized evidence avoid the environmental confounding seen by whole-tile transfer learning while preserving development-only evaluation?

Method🔗

• Private input: four approved, structurally masked development batches only.
• Spent and future holdouts: not transferred to the execution environment and not read.
• Evidence plan: ten Hero header/icon-rail proposals and eight Inventory-grid

proposals across the overlapping left and center tiles.

• Backbone: ImageNet-pretrained MobileNet V3 Small, frozen.
• Heads: independent multiple-instance linear classifiers trained from

sample-level labels; one matching localized region is sufficient evidence.

• Threshold: 0.65, fixed before evaluation.
• Evaluation: four leave-one-development-batch-out folds.
• Promotion gate, fixed before evaluation: at least 80% aggregate exact

accuracy, at least 70% exact accuracy in every fold, and no class above a 20% aggregate false-call-plus-miss rate.

• Accelerator: RTX 4070 through a non-preemptive accelerator lease.
• Internal execution record retained; identifier withheld.

Result🔗

Aggregate exact accuracy was 39/42 (92.9%). Hero's aggregate error rate was 4.8%; Inventory's was 2.4%. The frozen development gate passed. The job completed on CUDA in 35.6 seconds 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🔗

Accept localized panel evidence as the first candidate architecture to earn a fresh holdout. This does not promote a production model: the same small corpus still supplied both training batches and cross-batch evaluation environments.

Before reading new holdout pixels, train and durably package one final model from all approved development batches with the region plan, threshold, model metadata, and corpus digests bound together. Then collect a separately frozen, recognition-disabled live set with intentional Hero, Inventory, combined, and closed-panel state variety. Keep the operator's current ordinary gameplay out of the holdout unless it is explicitly admitted through that collection flow.