Technical Paper2026-04-14 Claim Extraction as a Neurosymbolic Primitive
LLM code generation fails because claims go unverified. This paper defines claim extraction as a neurosymbolic primitive: a neural layer extracts implicit claims from code, a deterministic oracle verifies each one. Under Kautz's taxonomy, this is Type 5 — Neural[Symbolic]. We benchmark the architecture on HumanEval and SWE-bench.