WorldSeed Inference: Type Safety for Claims Under Explicit Worlds

WorldSeed Inference v3 is a mathematically grounded claim-verification framework that treats scientific claims as programs evaluated inside explicitly declared worlds. Unlike conventional empirical AI evaluation pipelines, which implicitly embed assumptions about the world, sensing, and irreversibility inside datasets and preprocessing, WorldSeed makes these assumptions first-class semantic objects and enforces them mechanically. This release formalizes a minimal claim language equipped with: a semantic admissibility type system classifying claims into: Executable Bounded Unverifiable Blocked a small-step operational semantics for claim evaluation inside a world runtime, a formal proof of Progress and Preservation for executable claims, and a Type Safety theorem:well-typed executable claims do not go wrong and always reduce to a boolean verdict. The framework provides a foundation for building a deterministic claim linter (WorldSeed Claim-Lint), making semantic compliance and admissibility auditable, comparable, and mechanically checkable, independent of learning algorithms, model architectures, or benchmark scores. Key Contribution WorldSeed Inference v3 introduces a programming-language-level safety theorem for scientific claims under explicit world semantics: Executable claims are runtime-decidable and cannot get stuckunless the claim exceeds declared boundaries or invokes illegal actions. This makes inadmissibility categories (e.g., NOT COMPARABLE, BLOCKED) not a matter of opinion, but an operational consequence of declared world semantics. Scope This release does not propose new learning algorithms, training procedures, or empirical benchmarks. It strictly formalizes the semantic layer that precedes them: the admissibility and interpretability of claims under declared worlds.