Appearance

First Surgery ​

This walkthrough uses a synthetic weight matrix — no checkpoint or GPU needed.

Python API ​

python
import torch
from qriton_hlm import BasinSurgeon, compute_energy

# Create a synthetic Hopfield layer
d_model = 128
torch.manual_seed(42)
W = torch.randn(d_model, d_model) * 0.01
W = (W + W.T) / 2  # Symmetric

surgeon = BasinSurgeon.from_W(W)

# Survey: what basins exist?
survey = surgeon.survey(layer=0)
print(f"Basins found: {survey['num_basins']}")

# Inject a new basin
result = surgeon.inject(layer=0, seed=42, strength=0.1)
print(f"Existed before: {result['existed_before']}")
print(f"Exists after:   {result['exists_after']}")

# Verify it's there
v = surgeon.verify(layer=0, seed=42)
print(f"Is basin: {v['is_basin']} (cos={v['cos']:.4f})")

# Check the damage
diff = surgeon.diff(layer=0)
print(f"W change: {diff['relative_pct']:.2f}%")

# Undo
surgeon.restore(layer=0)
print("Restored to original.")

CLI ​

bash
$ qriton-hlm

hlm> load model.pt
hlm:model> survey 0
hlm:model> inject 0 42 0.1
hlm:model> verify 0 42
hlm:model> diff 0
hlm:model> restore 0

HLM Script ​

Save as first_surgery.hlm:

bash
load model.pt
survey 0
inject 0 42 0.1
verify 0 42
diff 0
restore 0

Run: qriton-hlm --script first_surgery.hlm

Next Steps ​