2026-04-27 - Phi-4-mini Run-Batches Landed and Layers 8–11 Passed

Intent: Record the run-batches orchestration outcome and the first bounded actual batch, following validation-before-scale discipline and the project ANE-only gate policy.

Setup: the Phi orchestration script now has a future-facing run-batches stage with --batch-size default 4, explicit --layer-end required to avoid silent full conversion, --stop-after-batches, resource preflight before each batch, and delegation to run-range for per-layer convert, compile, strict residency, and golden gates. Validation dry-run: run-batches --layer-start 8 --layer-end 16 --batch-size 4 --stop-after-batches 2 --gatekeeper-go --dry-run produced batches 8–12 and 12–16. Actual bounded run: run-batches --layer-start 8 --layer-end 12 --batch-size 4 --stop-after-batches 1 --gatekeeper-go.

Result: PASS. Layers 8–11 converted and compiled successfully. Strict residency passed for each layer: conv_total=4 conv_ane=4 conv_non_ane=0; compute_total=148 compute_ane=148 compute_non_ane=0. Golden passed: L8 cos=0.999868 rmse=0.014178 max_abs=0.085938; L9 cos=0.999853 rmse=0.015281 max_abs=0.056641; L10 cos=0.999875 rmse=0.016508 max_abs=0.109375; L11 cos=0.999807 rmse=0.016104 max_abs=0.093750.

Surprise / hurdle: The batch runner stayed bounded while requiring an explicit end layer and reusing run-range gates rather than duplicating per-layer logic.

Lesson: Batch automation is safe when it is explicit, stopped, preflighted per batch, and delegates every layer to the same residency and golden gates.

Next: No full 32-layer conversion, performance run, energy run, cleanup, or deletion was run; continue only through explicit gated bounded batches.

Refs: research/ANE_CHAIN_SCHEMA.md