2026-04-27 - Phi-4-mini-instruct Layer-0 Gate Residency Passed

Intent: Validate the smallest representative Phi-4-mini-instruct full-layer INT8 CoreML shard before scale-out, following the ANE residency gate and optimization discipline from the validation-first notes.

Setup: Layer 0 full-layer INT8 mlpackage conversion, CoreML compilation to mlmodelc, strict MLComputePlan residency check.

Result: Conversion succeeded; compiled mlmodelc succeeded; package/modelc size was 96M. Strict residency passed: conv_total=4 conv_ane=4 conv_non_ane=0; compute_total=148 compute_ane=148 compute_non_ane=0.

Surprise / hurdle: No fallback ops appeared in the strict plan for the representative layer-0 gate.

Lesson: Phi-4-mini-instruct layer-0 INT8 full-layer packaging is a viable ANE-resident pattern to consider for scale-out.

Next: Perf, energy, full conversion, and cleanup were not run; proceed only through the normal gated scale-out flow.

Refs: research/ANE_CHAIN_SCHEMA.md