episteme already computes turn classification, correction detection, and fact-type classification during extraction. None of that flows into the training record. The raw pair is captured while the labels the system generated on that turn are thrown away. For DPO/ORPO signal, add fields populated from episteme extraction outputs: turn classification, correction flag, conflict annotation, fact types. Requires schema_version bump.
episteme already computes turn classification, correction detection, and fact-type classification during extraction. None of that flows into the training record. The raw pair is captured while the labels the system generated on that turn are thrown away. For DPO/ORPO signal, add fields populated from episteme extraction outputs: turn classification, correction flag, conflict annotation, fact types. Requires schema_version bump.