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

Intent: After the layers 0–3 3-layer Phi-4-mini probe passed and the user said “ok, validate that,” validate the full 3-layer shard strategy across the whole 32-layer model, applying the validation-first notes call-hoisting/strength-reduction and whole-operation fusion discipline.

Setup: Planned non-destructive output directory: local artifacts. 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). For each range, compile if missing, then run strict MLComputePlan residency and range golden smoke.

Result: Intent recorded before execution; no new range artifact counts, placement, cosine, latency, energy, or perplexity numbers yet.

Surprise / hurdle: The first 288M 3-layer compile passed despite exceeding the older conservative shard-size caution, so every range must re-prove compile success, ANE residency, and numerical smoke instead of assuming scale-out safety.

Lesson: A fused-shard strategy is validated only when every planned range passes the same residency and quality gates, including the shorter tail.

Next: Run the full range validation only in the non-destructive probe directory; do not clean up/delete artifacts and do not run performance or energy benchmarking.

Refs: research/ANE_CHAIN_SCHEMA.md