Specification of Consequence-Based Learning

A learning procedure is specified in which the criterion of correctness is internal to the system and revisable, unlike prediction-based and reinforcement learning, where the criterion is external and fixed by the training signal. The system holds commitments, derives what they entail, evaluates the entailments against its other commitments, and revises when they cannot be sustained. The document states five necessary conditions, distinguishes the proposal from adjacent traditions, and formalizes the cognitive state as a product lattice and the self-state as a normalized measure over an anchor space, together with the derivation operator, stress map, choice functional, compressor, and coupled dynamics. Two limits are proved: the procedure can change by variation but not by inference to the best explanation, and its criterion has no external ground, so it can be coherent and wrong without detecting the discrepancy. Three substrates required for implementation do not currently exist.