Neural Recovery and Representation Diagnostics: Task Criteria and Prospective Monitoring

RT6 separates restoration of task performance from changes in hidden-representation diversity. It specifies damaged capability, repair intervention, measurement convention, and future evaluation target independently. A replacement experiment trains ten small ReLU classifiers, damages their output heads, and repairs those heads under a fixed budget. Four meet the local restoration tolerance while hidden activations remain unchanged across all three stages. Function-preserving reparameterizations also alter a representation-rank measure. These finite results challenge simple structural recovery thresholds rather than validate a general recovery composite. Earlier unauditable grokking claims are retired; the remaining proposal tests incremental monitoring value beyond ordinary performance controls.