Assistive Variant Curation in Clinical Cancer Whole-Genome Sequencing: Evidence Ablation and Time-Sealed Evaluation

We describe an assistive agent that curates driver variants in clinical cancer whole-genome sequencing at a two-site diagnostic laboratory and examine how evidence affects its adjudication. Five stages run ahead of a consensus conference of four trained reviewers, which finalizes every case and provides the reference labels used here. Two of the five stages call one frontier general-purpose multimodal model, in three calls, and no biology-specialized model is used at any stage; two more called a model earlier in development and now run on rules. On a development benchmark of 1,402 candidates, explicit predicates close 941 before any model is called, leaving 461 for adjudication. Withholding read-alignment and copy-number images raises the share returned to a reviewer without a verdict from 28.6% to 42.3%, withholding retrieval over past interpretations lowers specificity from 69.9% to 53.6%, and withholding the purity estimate moves no measure by more than 3.0 percentage points. Yet accuracy pooled over all 1,402 candidates remains within 94.2–94.8% in every arm, indicating that evidence can change which candidates the agent closes without a comparable change in the figure a concordance study would report. These measurements cover the verdict axis of reportability, not the clinical-significance assignment; they use a reference standard that shares provenance with the retrieval corpus and a release that preceded the deployed gate. We therefore pre-register OncoSEAL (Sealed Evaluation of Agents in the Loop): a retrospective set of over 1,000 cases sealed by time, a within-period concurrent control split by reader, an analysis plan hashed at the release tag and an external cohort with an independent reference. The seal is structural: deployment experience can re-enter only through an interface unavailable on sealed cases by construction. Because this paper precedes the first scoring, every number in it is a development-set or harness-contract measurement, and we make no claim about the agent's accuracy in deployment.