# Stored trained heads

`predict.py` uses the Python standard library and the fixed example `relu_head_example.json`. `heads/` contains all sixteen confirmation heads in the same exact-rational hinge format. Each is generated from its stored knot values, not retrained during packaging.

```bash
python models/predict.py -0.8 -0.2 0 0.2 0.8
```

In Python, use `load_model(path)` and `predict(x, model)` from `models/predict.py` to select another head. Rational segment checks establish head norm <=1, head Lipschitz <=2 and output Lipschitz <=1 for constant-tail extension over real inputs. Conditional error bounds concern only [-1,1] and use the reported floating numerical risk oracle; they are separate from exact regularity checks. The CLI is not the NumPy batch-throughput benchmark.