Skip to content

Real-World Acceptance Report

This report records the current S18 release-gate evidence. It is intentionally honest: default CI uses no network, opt-in external datasets run only from ignored local caches, and 1.0.0 remains blocked until the complete matrix, performance audit, and user commitment are finished.

Default No-Network Matrix

  • Iris: active multiclass native CPU softmax acceptance.
  • Digits: active flattened-image multiclass native CPU softmax acceptance.
  • Breast Cancer Wisconsin: active binary classification acceptance and sklearn CPU baseline record.
  • Diabetes: active regression acceptance, advanced objective acceptance, ranking smoke, and sklearn CPU baseline record.

Cached Opt-In Matrix

  • California Housing: active cached regression acceptance through sklearn fetcher.
  • Covertype subset: active cached large-row multiclass acceptance and real MPS parity smoke.
  • MNIST subset: active cached OpenML flattened-image multiclass acceptance.
  • Titanic: active cached OpenML missing-value and categorical workflow acceptance.
  • Adult Income: active cached OpenML categorical binary-classification acceptance.
  • HIGGS subset: active local-file large numeric binary-classification acceptance.

Degradation and Boundary Notes

  • Small datasets may run faster on CPU because GPU launch and synchronization overhead dominate.
  • IsolationForest and LearningToRankRegressor are CPU-suitable workflows; MPS requests warn and record the CPU backend decision.
  • MPS native multiclass softmax remains phased. CPU native softmax is the default native multiclass implementation; MPS multiclass uses observable OvR compatibility until native MPS softmax kernels are ready.
  • Optional external datasets are not downloaded during ordinary tests. Missing caches skip with copy-paste setup instructions.

Open Release Gates

  • S18.6/S18.6a: full model-quality, end-to-end performance, peak-memory, wheel-size, model-size, and permission audit.
  • S18.8: complete G17 matrix acceptance before planning 1.0.0.
  • S18.9/S18.10: user confirmation and formal PyPI 1.0.0 release.