Each learning event should identify its run, state revision, observation reference, prediction, feedback, complete candidate set, objective components, selected action, resource update, and successor state.
Separate fact classes
Store measured duration separately from estimated operation cost; endogenous resource separately from hard limits; phase classification separately from raw time-series values; and user feedback separately from ground-truth labels.
Replay is the test
A replay tool should rebuild structural and resource state from the ledger and either reproduce the same decisions or explain the source of nondeterminism through recorded seeds and environment data.