Withdrawing and Refusing Post-Deployment Learning Claims: An Instrumented Case Study of Prólogo, a Persistent LLM Assistant System

Claims that a deployed language-model agent has learned can fail in more than one way. An experience artifact may exist without being delivered to the model, a desired behavior may occur without a discriminating counterfactual, and a runtime may enforce a property that the model never learned. We report an instrumented longitudinal case study of Prólogo, a persistent LLM assistant system whose post-deployment learning claims were accepted, withdrawn, or refused as these evidence dimensions disagreed. We retain the historical withdrawal of negative efficacy interpretations; its original delivery finding is not independently established by the current reconstruction. Historical evaluations reported delivery as passed in two single-consumption same-task executions that also showed the pre-specified recovery behavior; neither observation provided causal attribution. A subsequent frozen no-lesson control and lesson-bearing treatment cohort was operationally valid yet non-discriminating: both arms recovered in 3/3 runs and the cohort was classified as inadmissible. Three sequential, adaptively designed control-only calibration episodes exposed distinct instrument defects. The terminal V3 episode had four structurally admissible runs, objective recovery in 1/4, bounded completion in 0/4, and recovery by sweep in 1/4; its frozen classification was cohort_unsuitable_sweep_dominant. We report no causal learning or transfer result. The contribution is a case-derived claim-analysis schema that separates artifact existence, release attribution, consumed-context delivery, behavior, counterfactual discriminability, recurrence, transfer, persistence, and deterministic enforcement. In the observed lineage, an evidence and adjudication structure and an operator-mediated governed process enabled earlier interpretations to be withdrawn or refused without rewriting their historical traces. External human evidence review remains outstanding.