Generative–Inverse Closure: A Unified Framework for Generation, Qualification and Recovery in Adaptive Dynamical Systems

Adaptive computational systems are often evaluated by forward generation, inverse diagnosis and recovery as separate tasks, allowing persistent but non-autonomous states, observer-compressed mechanisms or compensation-masked failures to be misclassified as valid effective states. We introduce Generative–Inverse Closure (GQR), a lineage-preserving framework in which generation, qualification and recovery form one conjunctive test. GQR combines explicit observation operators, non-compensatory admission axes, ablation–restoration–recovery evidence and typed closure debt. Archived computational studies expose distinct one-way failure modes. A new 64-seed single-lineage benchmark then generates relation-based coarse states, tests predictive and interventional qualification, applies controlled damage, restores the baseline dynamics and requires requalification after recovery. Generated candidates passed the declared qualification gate in 61/64 runs versus 4/64 randomized candidates, while 63/64 recovered candidates passed requalification. The result is methodological and model-level: GQR does not establish a universal physical law, but provides a falsifiable closure architecture for adaptive computational systems.