Predictions that remain connected to validation.

Moloracle treats model output as a hypothesis. Useful evidence includes held-out evaluation, versioned inputs, uncertainty, failure analysis, and a property-specific next calculation or experiment.