# Evidence: Convert PyTorch weights to Flax and locate numerical errors

Copy this template into your own portfolio and fill it after running your modified exercise.

## Evidence required

Keep source checkpoint and architecture identity, complete parameter mapping, per-layer max-absolute/relative-L2 errors, elementwise tolerances, input-gradient comparisons, wrong-epsilon diagnosis and a fresh-process converted-artifact receipt.

- Command: `python3 phases/15-deployment/08-pytorch-to-flax-numerical-parity/code/main.py`
- Working directory: repository root
- Python / JAX / NumPy versions: 
- Hardware and device: 
- My prediction before running: 
- My change to the exercise: 
- Exit code and meaningful output: 
- Why the result supports or contradicts my prediction: 
- Checkpoint result: 
- Artifact link or saved filename: 
- Portfolio status: saved in local workspace