Quick Start

Build an adaptive System 1 model, then choose how it decides, remembers, and adapts.

Build the System 1 contract

Wrap ARC 1, a trained classifier, or any score-producing callable in a grounded decision space.

system1.py

from gpbacay_arcane import DecisionSpace, System1def route(text, labels):    return {        "billing": 0.72 if "charged" in text.lower() else 0.10,        "shipping": 0.18,        "other": 0.10,    }model = System1(    predictor=route,    decisions=DecisionSpace(        ["billing", "shipping", "other"],        min_confidence=0.60,        min_margin=0.15,    ),    memory_weight=0.65,)# Learn a decided example immediately, without fine-tuning.model.remember("my card was charged twice", "billing")decision = model.decide("why was I charged again?")print(decision.label)       # billing, or None when ARCANE abstainsprint(decision.source)      # model+memoryprint(decision.evidence)    # recalled examples behind the decision

Advanced Usage with Resonance

Refine a current batch with isolated per-example targets, then optionally evaluate bounded adaptation.

adaptive-inference.py

from gpbacay_arcane import HierarchicalResonanceFoundationModel# Create an advanced model with biological neural principlesmodel = HierarchicalResonanceFoundationModel(    vocab_size=3000,    seq_len=32,    hidden_dim=128,    num_resonance_levels=4)model.build_model()model.compile_model(learning_rate=3e-4)# Train normally, including an explicit feedback-alignment step when desiredmodel.model.fit(X_train, y_train)model.train_feedback_step(X_train[:32], labels=y_train[:32])# Resonance changes latent state for this batch without changing learned weightspredictions = model.predict_with_resonance(X_test, resonance_cycles=4)# Optional test-time learning uses an objective, bounded updates, and rollbackadapted = model.adapt_at_inference(    X_test,    variables="all",    objective="entropy",    steps=2,    episodic=True,)

Sequence Models with PredictiveResonantLayer

For sequence or image-as-sequence tasks (e.g. MNIST), use PredictiveResonantLayer for local predictive resonance without hierarchical wiring.

predictive-resonance.py

from gpbacay_arcane import PredictiveResonantLayer, SpatioTemporalSummarization, BioplasticDenseLayer# Sequence of 28 steps (e.g. MNIST rows), 128 unitsx = PredictiveResonantLayer(    units=128,    resonance_cycles=3,    resonance_step_size=0.2,    return_sequences=True,    persist_alignment=False)(x)x = SpatioTemporalSummarization(d_model=128)(x)x = BioplasticDenseLayer(units=128, enable_inference_plasticity=True)(x)# then pool and classify

Tune resonance for the task: each cycle adds compute. Compare the same checkpoint at 0/1/2/4/8 cycles, then select the best balance of accuracy, calibration, latency, forgetting, order sensitivity, batch-size sensitivity, and distribution shift. See Inference-Time Adaptation for the full protocol.