cogniarc-nano-nn

CogniARC Nano-NN — experimental ARC classifiers

Small feed-forward networks trained with NumPy and inferable with Rust for experiments in the public CogniARC research harness.

They are comparison baselines, not general ARC solvers. The current CogniARC pipeline is rule-first and may use a micro-NN only as a hint or comparison. Low-confidence output must abstain or escalate to a separately verified path.

Bundled models

The historical accuracy values in older cards were measured on small or synthetic splits. They are not claims of ARC-AGI-3 generalization. Consult the source repository's dev/holdout reports and reproduce the relevant benchmark before drawing conclusions.

Where it is used

• CogniARC source
• micro_predictors.py
• Rules-vs-NN benchmark

Minimal JSON inference

import json
import numpy as np

with open("domain_classifier.json", encoding="utf-8") as stream:
    model = json.load(stream)

x = np.asarray([1.0, 1.0, 0.3, 0.35, 0.45, 0.02], dtype=float)
for index, (weights, bias, activation) in enumerate(zip(
    model["weights"], model["biases"], model["activations"]
)):
    matrix = np.asarray(weights).reshape(model["layers"][index + 1], model["layers"][index])
    x = x @ matrix.T + np.asarray(bias)
    if activation == "relu":
        x = np.maximum(0, x)
    elif activation == "sigmoid":
        x = 1 / (1 + np.exp(-x))
    elif activation == "softmax":
        shifted = x - np.max(x)
        x = np.exp(shifted) / np.exp(shifted).sum()

print(x)

Use the source feature extractors rather than inventing anonymous feature vectors. The example demonstrates the stored feed-forward format only.

Licence

Released under the MIT License.