Methods

NeuroTwin represents neural recordings as NeuralEventBatch objects and splits recording-level manifests before preprocessing or windowing. The experimental NFC path infers a latent neural field with shape [batch, time, nodes, latent_dim], evolves it causally, and compiles it into modality-specific observations. Pair-Operator is retained as an ablation for low-rank relational field updates.

Prepared training uses train for optimization, val for periodic evaluation and best-checkpoint selection, and test only for final held-out reporting. Reports expose selection_split, report_split, best_val_mse, and final test_* metrics so checkpoint selection cannot silently use test data.

Real-data v1 artifacts are prepared manifests, not raw public data. MOABB and BIDS adapters write leakage reports, manifest hashes, and benchmark metadata into run artifacts.