2026-04-27 - Phi-4-mini Four-Layer Fused Shard Residency Passed

Intent: Test whether a larger Phi-4-mini fused shard for layers [0,4) can remain ANE-resident beyond the prior 3-layer/288M probe, following the validation-first notes validation-before-scale and whole-operation fusion discipline.

Setup: Built non-destructively under local artifacts; converted and compiled one INT8 stateful CoreML shard for layers [0,4); ran strict MLComputePlan residency. Residency JSON: temporary output.

Result: PASS. Conversion and compile succeeded. Artifact sizes: .mlpackage 384M and .mlmodelc 385M. Strict residency passed: conv_total=16 conv_ane=16 conv_non_ane=0; compute_total=583 compute_ane=583 compute_non_ane=0; PASS=True.

Surprise / hurdle: A single 4-layer shard stayed ANE-resident despite exceeding both the prior 3-layer 288M probe and the conservative ~250M compiled-shard guidance.

Lesson: Phi-4-mini fused-shard size limits are empirical and range-specific; exceeding conservative size guidance does not imply ANE fallback when strict residency still passes.

Next: Run golden validation for [0,4), then validate representative/all 4-layer ranges before any runtime migration. No cleanup/deletion and no energy benchmark were performed.

Refs: research/ANE_CHAIN_SCHEMA.md