2026-04-27 - Phi-4-mini Four-Layer Batch Runner Intent
Intent: Add a future-facing Phi-4-mini batched build runner that advances only bounded four-layer batches, following validation-before-scale discipline and the project ANE-only gate policy.
Setup: Planned orchestration: process one explicit four-layer batch at a time, check resources between batches, and reuse existing per-layer gates: convert, compile, strict MLComputePlan residency, and numerical smoke.
Result: Intent recorded before implementation; no artifacts, placement numbers, latency, energy, cosine, or perplexity results yet.
Surprise / hurdle: The runner must make batched progress convenient without silently running all 32 layers or bypassing per-layer gates.
Lesson: Batch automation is safe only when batch size is bounded, resource checks happen between batches, and every layer still passes the same gates.
Next: Implement the runner non-destructively; do not run performance or energy tests, do not clean up/delete artifacts, and require explicit user action for any full 32-layer scale-out.
Refs: research/ANE_CHAIN_SCHEMA.md