MechTrace: Relation-Complete Evidence Weaves for Mechanistic Biomedical QA

A mechanistic biomedical answer can name the correct drug, target, and outcome while leaving their relation unsupported, substituting endpoint plausibility for inspection of the mechanism. MechTrace makes requested relations inspectable with a connected evidence weave of typed edges, aligned snippets, provenance, and stable edge–snippet identifiers. The weave is both generator context and legal citation set; pre- and post-decoding checks expose the first relation without a valid witness. A path-matched comparison fixes upstream path discovery, generator, evidence snapshot, and retrieval budget, then contrasts this closed interface package with TypedPath-Open. On 600 fixed questions, MechTrace raises all-question correctness from 58.2% to 63.0% and reduces unsupported-relation escapes from 69 to 41, while coverage changes by 0.5 points. Among answered outputs, citation correctness improves by +2.5 points on BioASQ-mech and coherence by +2.7 points on DrugMechQA. The citation advantage grows from +0.8 to +4.6 points with BioASQ-mech link demand, persists with Qwen2.5-72B-Instruct, and accompanies 78.0% versus 49.0% detection when a required bridge is removed. These decision-level gains and missing-edge signatures demonstrate the verification value of relation-complete evidence weaves.