2026-04-27 - Phi-4-mini Layers 2–3 Bounded Smoke Passed

Intent: Record the bounded layers 2–3 Phi-4-mini-instruct guarded outcome before any scale-out, following the established quality-before-performance workflow and validation discipline.

Setup: Ran only layers 2 and 3 through INT8 full-layer mlpackage conversion, compile, strict MLComputePlan residency, and per-layer PyTorch-vs-CoreML numerical smoke. temporary JSON outputs.

Result: PASS. Layers 2 and 3 INT8 full-layer mlpackage conversion succeeded; compile succeeded; each mlpackage/mlmodelc was 96M. Strict residency passed for both: conv_total=4 conv_ane=4 conv_non_ane=0; compute_total=148 compute_ane=148 compute_non_ane=0. Numerical smoke passed: layer 2 cos(hidden)=0.999893, rmse=0.008980, max_abs=0.058594; layer 3 cos(hidden)=0.999878, rmse=0.010047, max_abs=0.090942.

Surprise / hurdle: The two-layer batch stayed bounded and preserved strict ANE residency with no observed non-ANE compute ops.

Lesson: Phi-4-mini INT8 full-layer conversion remains repeatably ANE-resident and numerically close through layers 2 and 3.

Next: No full conversion, perf, energy, or cleanup was run; proceed only through the gated scale-out flow.

Refs: research/ANE_CHAIN_SCHEMA.md