"""Independent NumPy scalar dynamics; never calls the tensor step.""" import numpy as np def scalar_step(model, observation, state): """Return (logits, next_state) using a snapshot of model scalar parameters.""" weights = {name: tensor.detach().cpu().numpy().copy() for name, tensor in model.state_dict().items()} observation, state = np.asarray(observation), np.asarray(state) c = model.config if observation.ndim != 2 or observation.shape[1] != c.input_dimension: raise ValueError('invalid observation shape') if state.shape != (observation.shape[0], c.neuron_count, c.state_dimension): raise ValueError('invalid state shape') aggregate = np.zeros_like(state) for edge, (source, target) in enumerate(zip(weights['scalar_sources'], weights['scalar_targets'])): aggregate[:, target] += (state[:, source] * (weights['edge_weights'][edge] / weights['scalar_degree'][target])) sensory = observation @ weights['input_encoder.weight'].T + weights['input_encoder.bias'] drive = (state * weights['self_gain'] + aggregate @ weights['local_transform'] + sensory[:, None] * weights['input_profile'] + weights['neuron_bias']) next_state = np.tanh(drive) readout = next_state[:, weights['readout_neurons']].mean(axis=1) logits = readout @ weights['output_decoder.weight'].T + weights['output_decoder.bias'] return logits, next_state