2026-04-29 - Phi Dead Artifact Cleanup Outcome

Intent: Reclaim disk by deleting only the approved Phi dead artifacts already ruled out by CPU fallback, slower profiling, or duplication, following measurement discipline.

Setup: Deleted local artifacts; generated top-k LM-head top1 s0 .mlmodelc/.mlpackage; generated 3-way and 8-way LM-head .mlmodelc/.mlpackage artifacts while preserving their manifests; and slower 5-layer tail generated artifacts [20,25), [25,30), plus duplicate [30,32).

Result: Cleanup completed. Disk free increased from about 6.2G to 14G; du reported the deletion set at 9.3G total. The current 20+4+6+2 baseline artifacts and the batch-4 LM-head set were preserved.

Surprise / hurdle: The main risk was avoiding accidental deletion of useful manifests or current baselines while removing large generated experiment outputs.

Lesson: Destructive artifact cleanup is safe when deletion targets are tied to recorded failure/slower evidence and protected baselines are named explicitly.

Next: Continue from the preserved 20+4+6+2 baseline and batch-4 LM-head artifacts; require separate approval for any further artifact deletion.

Refs: research/ANE_CHAIN_SCHEMA.md