Module claim_support 

The support arithmetic: how strongly a claim triple is backed by a set of cited fact quins. Pure over NQuins — no I/O, fully testable.

A claim is the triple (subject, predicate, object) of AgentOutput::semantic_quin. Two complementary, real signals are computed over the cited facts:

• Role support — the strongest single fact that matches the claim’s components in the same role: an exact triple match scores 1.0; a predicate+object or subject+predicate match 2/3; a single role 1/3. This is “is this exact relation attested?”.
• Entity grounding — the fraction of the claim’s two endpoints (subject, object) that appear anywhere in the evidence. This catches a claim that is a legitimate multi-hop consequence of cited facts even when no single fact is the whole triple: “are the things I’m talking about even in the evidence?”.

The combined score takes the stronger of role support and a capped half-weight of entity grounding, so an exact attestation always dominates while a claim whose endpoints are at least cited gets partial (review-band) credit.

Structs

GroundingReport

Per-claim grounding signals plus the combined score in [0, 1].

Functions

component_support

Strongest same-role support for claim across facts (max role overlap). 0.0 for empty evidence.

entity_grounding

Fraction of the claim’s endpoints {subject, object} that occur in any role (subject, predicate or object) of any cited fact.

report

Build the full grounding report for claim over facts.