A deterministic, auditable engine for cross-theory claim comparison: cross-field discovery with built-in verification

I present a deterministic engine that compares scientific claims across competing theories. Each claim is first reduced to a fixed schema of typed fields, in the style of information extraction. The fields are its entity type, its axis (the aspect of the system it addresses), its polarity (assertion versus negation), and a value. A deterministic rule-based verifier, which I call the judge, then compares two claims from these fields alone. Its output is a verdict: either a grade of correspondence, which says how strongly the two claims agree, or a fork, a question on which the theories make incompatible commitments. The judge is reproducible, provable and auditable, and I re-express it three ways that agree with it exactly. As a logic program, verdict equivalence is 1.0 (n = 213 pairs). As a first-order theory under an SMT solver, congruence is 1.0 (n = 311 verdicts). As machine-checked proof objects, coverage is 0.9842 (n = 316), with 0 integrity violations (n = 316). On the discovery side, the engine recovers the COGITATE adversarial-collaboration key at 2 of 3 predictions (revised; n = 3; not significant against a 1000-key permutation null, p = 0.075). The eLife synergistic-workspace key recovery is below chance: 0 exact / 6 partial of 14 validated rows, with empirical p = 0.981 (n = 1000; random keys recover more). The method's limits are first-class results. Both key recoveries carry their chance rates, single-pass extraction carries a measured noise floor (its test-retest reliability across repeated runs), and the extraction front-end under-propagates negation.