EvidenceWeave: Compiling Time-Bounded Evidence into Inspectable Mechanisms and Discriminative Tests

One omitted contradiction can reverse the rescue experiment chosen from an otherwise well-formed mechanism graph. EvidenceWeave keeps that failure visible by compiling the path from each causal edge to its source span, accepted or uncertain state, unresolved prerequisite, rival mechanism, and outcome-separated test. The accompanying benchmark contains 312 CRISPR perturbation cases and 4,210 expert-identified support or contradiction opportunities, including 715 contradictions, from literature published before January 2022. Chronology is enforced before retrieval ranking: fixed-volume future-document injection yields 0.089 leakage and 8.3% support inflation at 20%, whereas the date-filtered pipeline has zero detected leakage. Against a typed Llama-3.1-70B control matched on training objects, retrieval, schema, and decoding, EvidenceWeave lowers unsupported-edge rate from 0.246 to 0.142 while raising contradiction coverage from 0.302 to 0.478 and evidence-opportunity recall from 0.601 to 0.672; 299 additional annotated opportunities reach reviewers with similar accepted-edge mass. A separately blinded 72-case annotation panel preserves the edge ordering, while a 48-case human target yields written-outcome-separation AUC of 0.706 versus 0.584 for the control. Counterbalanced review also falls from 11.1 to 8.7 minutes, with faster review in 49 of 60 pairs. The resulting model-selection criterion rewards evidence exposure and adverse-evidence recovery only when asserted edges remain supportable and the linked object reduces expert reconstruction.