GPU Driver v1.0.0-rc2 — out-of-time baseline gate

GPU Driver v1.0.0-rc2

This prerelease hardens the Python DirectML inference driver with an executable
out-of-time naive-baseline publication gate.

Added

• gpu_forecast_driver.py validate
• SHA-256-bound inference of the historical training cutoff from
  lstm_predictions.json
• Recursive post-cutoff evaluation of every eligible exported ONNX graph
• Aggregate and per-commodity model-versus-persistence MAPE, MAE, RMSE, and R²
• A typed prediction publication_gate
• A fail-closed error that identifies the project-pinned DirectML interpreter
  when another active Python environment exposes only CPU/Azure providers

Current measured result

The current exported checkpoints failed the new gate on observations after the
historical forecast origin:

• evaluation window: February–June 2026
• eligible models: 59
• observations: 295
• LSTM MAPE: 85.6945%
• naive-persistence MAPE: 7.5142%
• LSTM MAE: 141.9462
• naive-persistence MAE: 8.4031
• publication status: withheld_failed_validation

The DirectML rerun still produced 1,062 local experimental values from 59
current models and recorded 13 stale/incomplete model series. Those values are
not release assets and are not public forecast evidence.

Unchanged execution evidence

The previously released node-profile and latency benchmark remains applicable
to the same exported Rice LSTM graph:

• 15 of 18 profiled node events used DmlExecutionProvider
• both LSTM nodes used DirectML
• three nodes used CPU fallback
• GPU/CPU maximum absolute output difference: 0.00000381
• CPU median latency remained lower at every measured batch

DirectML placement is not called acceleration. This is an ML execution driver,
not a Windows display or kernel driver.

Release assets

• benchmark_current.json
• benchmark_latency.svg
• export_manifest.json
• validation_current.json

Predictions and model bundles are intentionally excluded.