Structural Underdetermination in AI Alignment Claims: A Preregistered Theoretical Benchmark (Layers 1–3)

This Zenodo archive contains the frozen theoretical core (Layers 1–3) of a five-layer research program investigating structural underdetermination in AI alignment research. It establishes a formal necessity result: any system claiming reliable coherence under open-domain novelty—defined as preservation of alignment-relevant properties under non-stationary tasks, recursive self-modification, adversarial pressure, and evaluator reformulation—must explicitly instantiate a minimal property bundle (P) consisting of: Closure under composition, Recoverability, Explicit invariants, Evaluator representability (internal encoding of the evaluator projection ΠeΠe), and Stability under recursion. The archive includes precise definitions of the target regime (R) and property bundle (P), a full proof of the Necessity Theorem, and supporting lemmas (Certification Barrier, Adversarial Lower Bound, Evaluator Encoding Requirement). All content is frozen as of January 17, 2026, and will not be modified prior to, during, or after the execution of the independent empirical literature audit (Layer 4). This benchmark enables a non-circular, reproducible audit of representational explicitness in the AI alignment literature, separating theoretical foundations from empirical analysis to mitigate hindsight bias and theory-data circularity.