MIB-V15: A Unified Certification Protocol for Causal Logical Encoding in Neural Networks

We present MIB-V15, a unified certification protocol for determining whether a neuralnetwork causally encodes logical propositions rather than merely simulating them throughsurface-level heuristics. The protocol integrates five sequentially blocking steps —QK-Causal Tracing, Pareto-Alignment via the Kneedle algorithm, Divergence-Adjusted IIA(DA-IIA) normalised by Earth Mover Distance, MIB-SAT structural invariance acrossseven benchmark families, and AI-SMT formal verification via Marabou — into a coherentcertification pipeline applicable to both neural SAT solvers (Track A) and large languagemodels producing Chain-of-Thought reasoning (Track B). A cross-track arbitrationframework resolves contradictions between internal mechanisms and textualself-explanations. The protocol outputs one of six governance decisions (Scenarios A–F),providing actionable deployment guidance. All thresholds are empirically calibrated onsymbolic solver baselines (MiniSat, Kissat) to eliminate arbitrary parameter choices. Thisdocument constitutes the formal specification of the pipeline, its theoretical justification,and its implementation architecture, as generated by the meta-forge automated codegeneration system.