# Trained PyTorch weights

`PyTorch/DT` contains **15 original checkpoints**: E1/E2/E3 for C_5, C_10, S_100, SC_5, and SC_10. `PyTorch/BC` contains **3 original BC checkpoints**. Original file names and bytes are preserved.

The DT originals were recovered from `DT Models.zip` at repository revision `b68573995aea82d97a599ee7ea0c9b2ea8d88f8b`. The original ZIP hash is `7d8be2d43a5a8b6caac30cf11fafb5d5a9520e07c5de843b410245bad7a463de`. Its 15 paired ONNX members each have exactly the same SHA-256 as the corresponding evaluated export. Existing local E1_C5 and E3_S100 weights also matched the archived originals. The three BC weights were previously recovered from the same pinned revision and were copied from the verified local originals.

This confirms source-archive pairing and file integrity; no new inference, export, or training was performed during packaging. Existing numerical parity evidence covers the explicitly reported representative checkpoints, not every model.

PPO's exact evaluated trained policy is included at `../05_ONNX_Models/PPO/V12 PPO.onnx`. A matching native V12 training checkpoint was not located. Older V2/V10 `.pt` files in the Unity source assets are not substitutes for V12.

See `../08_Manifests/evaluated-models.json` and `dt-model-original-provenance.json` for per-file links and hashes.