Constrained Audit Layer

In two-model ghost-optic and bilayer cases, deterministic max-disagreement performed as well as or better than the LLM under identical constraints. Audit ablations indicate that constrained falsification, not LLM use itself, is the active ingredient. The LLM audit layer was tested only under bounded menus and strict guardrails rather than as an unconstrained scientific reasoner. The audit layer can only choose between log-GP active mode and physics mode and nominate pre-vetted audit points. The strategic audit layer is implemented as a constrained LLM committee. The audit layer cannot fit Hamiltonians, update likelihoods, alter posteriors, bypass kinematic checks, or emit arbitrary coordinates.