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

Intent: Continue Phi-4-mini-instruct ANE support with a bounded batch of layers 2 and 3 only, using the established guardrails and validation-before-scale discipline.

Setup: Planned flow: guarded preflight → convert → compile → strict MLComputePlan residency → per-layer PyTorch-vs-CoreML numerical smoke for layers 2 and 3. Constraints: no full 32-layer conversion, no perf/energy run, and no cleanup or deletion.

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

Surprise / hurdle: The immediate risk is keeping a two-layer batch bounded so it exercises repeatability without becoming scale-out.

Lesson: Small bounded batches can test process repeatability while preserving the ANE residency and quality gates before any expensive expansion.

Next: Run only layers 2 and 3 through the guarded flow; record residency and numerical-smoke results in a follow-up entry.

Refs: research/ANE_CHAIN_SCHEMA.md