When Valid Evidence Goes Unused in Oncology Agents

Rejecting an ineligible oncology citation can leave an evidence inventory unfinished even when eligible alternatives remain available. We study this failure under record-specific molecular, disease, and origin conditions, separating evidence exposure, valid submission, and inventory recall. With one model and constructed rejected proposals, a fixed-exposure study supplies two eligible alternatives in every ten-record payload, yet diagnostic repair ends in abstention in 53/64 co-alteration trajectories. A follow-up using a common clarified prompt compares fresh answering with supplied rejection context on the same cases and candidates: fresh answering yields 54/64 valid inventories versus 43/64 for diagnostic repair, a difference of +17.2 percentage points ($95\%$ paired variant-cluster interval [2.8, 33.3]). Reframing the proposal and feedback as user review material does not consistently restore fresh yield. Separately, a post hoc replay on saved search paths raises inventory recall from $60.3\%$ to $67.2\%$ through deterministic serialization of acquired eligible records. These findings support evaluating recovery after rejection, preserving the eligible inventory, and assessing biological interpretation as distinct workflow objectives.