2026-04-27 - Phi-4-mini-instruct ANE Support Scaffolding Intent
Intent: Start Phi-4-mini-instruct ANE support with safe scaffolding only: reusable analyzer/preflight and orchestration scripts before any expensive conversion. The plan follows the ANE-only mandate, quality-before-performance gating, and optimization discipline from the validation-first notes (measure and validate before scaling an implementation).
Setup: Workspace: this repo; model seed artifact: the local Phi-4-mini GGUF weights; planned baseline: INT8 per-tensor CoreML shards targeting ANE. Initial implementation scope is non-destructive preflight/analyzer/orchestration code only, with disk/RAM/cache guardrails and no full conversion, no cleanup of model/output artifacts, and no benchmarking.
Result: Intent recorded before implementation. No artifacts produced yet; no residency, latency, energy, cosine, or perplexity numbers yet.
Surprise / hurdle: Phi-4 support must be structured so that scaffolding cannot accidentally trigger heavyweight conversion or destructive cleanup while still encoding mandatory gates.
Lesson: New model support should begin with guardrailed orchestration that makes ANE residency and golden quality gates unavoidable before any performance work.
Next: Implement the analyzer/preflight and orchestration scripts; require MLComputePlan residency validation plus golden quality validation before any benchmark or scale-out conversion.
Refs: research/ANE_CHAIN_SCHEMA.md