Deterministic Verification for Cross Domain Scientific Claim Comparison

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

When researchers compare theories across disciplines, they need to know whether two frameworks commit to the same structural relationships. Not whether they sound alike. Existing tools cannot make this distinction. Embedding based systems produce false positives because cosine similarity and structural correspondence are mathematically different relations. We describe a neurosymbolic system built to separate generation from judgment. Language models extract typed claims under schema constraints and a cross model agreement gate. All judgment is deterministic code: VF2 subgraph isomorphism and maximum common subgraph for matching, a four rung grading ladder with explicit ceilings, and a versioned instrument capability map that assesses whether each disagreement is testable with current instruments. No model gets the last word. Same inputs, same output, verifiable by content hash. Two results. First, we built and deployed a three level embedding system before this one. It produced 6,341 mappings. A structural audit found systematic false positives where vocabulary overlapped without structural correspondence, and systematic false negatives where structural correspondence existed without shared vocabulary. We killed it. Second, the rebuilt system, blind tested against COGITATE (11 laboratories, 19 months), recovered the same structural disagreements the experts found and surfaced one they did not resolve. None of the algorithms are new. VF2 is from 2004. What did not exist was the typed schema they operate on and the authored assets they require. The contribution is the assembly, the evaluation, and the demonstration that the generator verifier pattern extends from mathematics and code to empirical science.