2026-04-27 - Phi-4-mini Full 3-Layer Shard Strategy Validated

Intent: Record the completed full-model validation of the Phi-4-mini 3-layer fused-shard strategy, applying the validation-first notes call-hoisting/strength-reduction and whole-operation fusion discipline.

Setup: Validated ranges [0,3), [3,6), [6,9), [9,12), [12,15), [15,18), [18,21), [21,24), [24,27), [27,30), and tail [30,32) in local artifacts; no cleanup/deletion and no energy benchmarking.

Result: PASS. All 11 compiled .mlmodelc artifacts are present; total directory size is 6.0G. Strict residency passed all ranges: 3-layer ranges conv=12/12/0 and compute=438/438/0; tail conv=8/8/0 and compute=293/293/0. Golden smoke passed all ranges: cos_hidden min=0.99952924 mean=0.99964909 max=0.99976834; max RMSE=0.09616414; max_abs=0.47656250.

Surprise / hurdle: Every planned fused range, including the 2-layer tail, compiled and remained ANE-resident despite the earlier shard-size caution from the first 288M 3-layer probe.

Lesson: The 3-layer Phi-4-mini fused strategy is validated across the full model for ANE residency and numerical smoke, so the next bottleneck is runtime migration rather than more per-range proof.

Next: Proceed to fused-shard manifest/runtime migration if desired; do not run energy benchmarking unless explicitly requested.

Refs: research/ANE_CHAIN_SCHEMA.md