varied Inventory position transfer

Labnote 013: varied Inventory position transfer🔗

Question🔗

Does adding new Inventory-only development batches staged at left, right, and center positions repair the localized candidate's position dependence?

Method🔗

• Private input: seven approved, structurally masked development batches.
• New data: three four-sample Inventory-only batches staged at left, right, and

center positions.

• Spent holdout: excluded and not read.
• Model, regions, threshold, and development gate: unchanged from Labnote 011.
• Evaluation: seven leave-one-development-batch-out folds.
• Accelerator: RTX 4070 through a non-preemptive accelerator lease.
• Internal execution record retained; identifier withheld.

Result🔗

Aggregate exact accuracy was 48/54 (88.9%). Inventory generalized across all three new position folds, but the second original environment collapsed through four missed Hero states. The 44.4% fold violated the frozen 70% floor, so the development gate failed and no candidate was emitted.

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🔗

Do not package a successor candidate. The new data repaired the targeted Inventory position behavior in development, but exposed order-sensitive Hero head instability after the negative-class distribution changed.

Replace per-sample optimizer steps with a deterministic, class-balanced full-batch multiple-instance objective. Re-run the same development-only folds under the existing gate. Do not read or rerun the spent holdout.