workload-validation.md

Native integration succeeds; added search value is unproven

Measured on 2026-10-14 on Apple M1 MPS with CPU fallback disabled. The runner, tests and prospective protocol were frozen at ea94b31 before execution. All 50 tests passed before the run.

The workload contains two RMSNorm layers of width 512. It is a synthetic micro-workload, not a pretrained transformer. Search sees two discovery shapes, then freezes its selection before measuring two new confirmation shapes. Each timing uses five randomized-order repetitions and synchronized execution.

Seed	Chosen policy	Confirmation speedup vs original	Vs always-native
142	Native RMSNorm	5.262891×	1.000000×
143	Native RMSNorm	4.804880×	1.000000×
144	Native RMSNorm	4.954127×	1.000000×
145	Native RMSNorm	4.959754×	1.000000×
146	Native RMSNorm	4.937313×	1.000000×

Speedups are equal-weight geometric means across the two confirmation shapes, computed from each method's median timing. All 480 CPU-reference correctness comparisons passed, including zero, small and large values and strided inputs. All 30 identical-native-code controls were rejected. No confirmation case regressed against either comparison baseline.

The frozen engineering gate passed in 5/5 runs. The added-search-value gate failed, with 0/5 runs passing: the selected policy is the always-native comparator itself. This demonstrates correct native integration and repeatable workload improvement against explicit operations. It provides no evidence that this bounded search improves on simply choosing native RMSNorm. Native substitution is an established optimization, not a novel kernel result.

The compressed raw report contains sources, hashes, every timing order/block, correctness records, selection decisions, control trials and machine-evaluated gates. Uncompressed SHA256: 83c549d3fdd42fe725ad2248c78d2811b5c042f8d06eabdcb89fdd6d77e89c3c. The aggregates were separately recomputed from the raw timing blocks.

To reproduce from a clean committed source checkout:

PYTORCH_ENABLE_MPS_FALLBACK=0 python -m experiments.workload_validation \
  --device mps --output workspace/workload-confirmation

Use --device cpu for a separate CPU experiment. No CUDA, backward/training, mixed-precision, pretrained-model latency or broad workload advantage follows from these MPS measurements. Shared shapes and a single machine limit the scope; five seeds and 30 controls are descriptive checks, not significance tests or a measured general false-promotion rate.