Bidirectional Mechanism Discovery and Structured Gap Localization

We integrate a general baseline-preserving generative–inverse workflow with a structured method for localizing mechanism gaps. The workflow preserves a versioned baseline, infers candidate mechanisms from observations, replays candidates through declared forward and observation operators, and represents residual discrepancies by mechanism classes rather than a single scalar error. The combined framework separates candidate generation, non-identifiability analysis, active discrimination, and causal verification. A structured gap is a diagnostic hypothesis, not proof that a mechanism is causally necessary; ablation, restoration, and recovery remain distinct tests. The contribution is methodological. It does not establish unique mechanism recovery from aggregate observations or validate a specific physical generator.