lcruntime

Python binding for Local Coherence — a CPU-native runtime for deterministic local-propagation inference.

Wraps liblc.so / liblc.dll via ctypes. No PyTorch, no ONNX, no GPU.

git clone https://github.com/bracoTuxbr/lcruntime-python
cd lcruntime-python
pip install -e .

You also need the compiled liblc.so (Linux) or liblc.dll (Windows) from the main repo. Point at it via env var:

export LC_LIB=/path/to/liblc.so

(pip-installable wheel with bundled native lib is on the roadmap.)

import lcruntime as lc
import numpy as np
t = lc.Tissue1D(1024)
t.set_kernel(lc.KERNEL_CANONICAL)
t.set_sig_delta(4)
t.inject(512, 1, 30000)
t.step(100)
field = t.get_field()
print(f"peak={field.max()} active={int((field > 0).sum())}")

t = lc.Tissue1D(n_cells)
t = lc.Tissue2D(rows, cols)
t.set_kernel(lc.KERNEL_CANONICAL | lc.KERNEL_SIMPLE_AVG | lc.KERNEL_EMA)
t.set_sig_delta(n) # 1 = bit-exact mode, 4 = golden default
t.set_dirty_enabled(True) # enable/disable freeze
t.inject(pos, count, magnitude)
t.step(n_generations)
field = t.get_field() # numpy view (uint16)
t.close()

✓ You want to integrate LC into a Python pipeline (audio, time-series, network telemetry) ✓ You need bit-exact reproducibility across machines ✓ Your deployment target is CPU-only / sub-watt edge

✗ You need a pre-trained model. LC is not a model — it's a runtime. ✗ You need state-of-the-art accuracy on a standard ML benchmark. LC trades accuracy for determinism + speed in specific niches.

See the main repo for the paradigm explanation, the paper, and an honest evaluation of where LC wins and loses.

Apache 2.0 — see LICENSE.