2026-04-27 - Phi-4-mini Layers 4–7 Bounded Run-Range Passed

Intent: Record the actual bounded Phi-4-mini layers 4–7 build+test outcome, following validation-before-scale discipline and the project ANE-only gate policy.

Setup: Ran the bounded layer-range stage of the Phi orchestration script; per-layer flow: preflight, convert, compile, strict MLComputePlan residency, and numerical smoke.

Result: PASS. Layers 4, 5, 6, and 7 converted and compiled successfully. For each layer, mlpackage=96M, mlmodelc=96M, and meta=4.0K. Strict MLComputePlan 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. Numerical smoke passed: L4 cos=0.999892 rmse=0.012303 max_abs=0.044922; L5 cos=0.999893 rmse=0.012867 max_abs=0.101562; L6 cos=0.999875 rmse=0.012494 max_abs=0.058105; L7 cos=0.999883 rmse=0.012904 max_abs=0.042969.

Surprise / hurdle: The first real four-layer run-range stayed bounded while preserving strict ANE residency and numerical-smoke gates for every layer.

Lesson: The bounded run-range runner can advance four Phi-4-mini INT8 full-layer shards at a time without observed ANE fallback or numerical-smoke regression.

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

Refs: research/ANE_CHAIN_SCHEMA.md