Native Arm64 / controlled before-and-after evidence

Half the model memory. 1.73× faster. Prediction parity preserved.

The deadline-era optimization converts the same forecast workload from FP64 to FP32 and measures both artifacts on a native Arm Neoverse-N2 runner. The evidence receipt binds hardware, model hashes, workload, latency, size, and quality.

Native Arm64 evidenceAEOLUS / benchmarks

1.73×Median speed-up738.349 → 426.358 µs

50%Array memory reduction28.76 MB → 14.38 MB

47.1%Artifact size reduction2.13 MB → 1.12 MB

4.88e-6Maximum normalized driftPassed 1e-4 parity gate

ArtifactFile sizeMedianp95

FP64 baseline2,126,337 B738.349 µs797.245 µs

FP32 optimized1,124,273 B426.358 µs445.057 µs

Prediction parity passedMaximum normalized drift 4.879×10⁻⁶ against a predeclared 1×10⁻⁴ acceptance threshold.