Human Correction Is Not Retroactive: Causal Closure and Operational Repair in Longitudinal Human–AI Systems

Longitudinal human–AI systems often treat a human correction as one more event appended to an interaction history. Yet the textual presence of a correction does not ensure that interpretations, inferences, decisions, memories, or actions derived from the corrected representation lose operational validity. This paper distinguishes three levels of correction: propositional correction, in which the agent locally accepts the revised formulation; genealogical correction, in which it identifies causally dependent descendants of the invalidated representation; and operational correction, in which downstream effects are revised, re-executed, repaired, contained, or escalated. We model a longitudinal trajectory through provenance, dependency, validity, and effects, and define correctional causal closure as the explicit treatment of all materially affected descendants without destroying independently supported states or erasing the genealogy of the error. The proposal is situated relative to belief revision, conversational memory, context pollution, in-place feedback, and editable transcripts. A longitudinal occurrence in the IH-001 system is used only to make the problem observable, not as general empirical validation. We propose a correction-aware system architecture, six testable propositions, an experimental protocol, and metrics for genealogical recall, genealogical precision, residual contamination, valid-state preservation, and operational repair. The minimal claim is that human authority to correct does not guarantee that an artificial agent has a general and reliable competence to calculate and execute the retroactive scope of that correction.