2026-04-28 - Phi Public Two-Call Reference and Stream-Level Event Probe

Intent: Prove whether the stream-level Phi stage-B zero was caused by bad synthetic inputs or by the unsupported two-operation stream-level stream path.

Setup: Added local artifacts, a public-only CoreML reference that loads phi4mini_layer16_24_q8.mlmodelc and phi4mini_layer24_30_q8.mlmodelc in a fresh process. It runs public stateful prediction for stage A, feeds A’s hidden to stage B, and avoids mixing public prediction with unsupported operation-pool mutation.

Result: Public stage A exactly matches the stream-level stage A for the synthetic Phi input: sample [1.498046875, 5.98828125, 4.94921875, -2.49609375, ...], sum -337.912079. Public stage B is nonzero: sample [-8.0625, -0.251953125, -0.564453125, -5.12890625, ...], sum -166.729431. The stream-level path still returns all-zero stage B, so the remaining issue is not input generation; it is a multi-operation dependency, event, or output synchronization problem.

Surprise / hurdle: Disassembly confirmed plausible stream-level event signatures. A guarded experiment retains stage A’s completion event and binds it as a dependent event on stage B; both manual calls return 0, but the process segfaults during/after the second operation prepare path. The experiment is hidden behind --bind-stream-events and is not the default path.

Lesson: The event API is real, but CoreML’s event lifecycle/order matters. Directly attaching the retained completion event is too early, too late, or missing future-value bookkeeping.

Next: Reproduce CoreML’s event lifecycle at a higher level before making any correctness or performance claim.

Refs: research/ANE_CHAIN_SCHEMA.md