When Valid Evidence Goes Unused in Oncology Agents

Large language model (LLM) agents can avoid invalid claims yet leave eligible evidence unreported. We investigate two forms of this failure in oncology evidence inventories: abstention despite visible eligible alternatives and omission of eligible records already retrieved. We use constructed somatic lung-cancer queries, with eligibility defined by encoded molecular, disease, and origin rules. In a retrospective synthesis of 56 cases and 496 runs, we compare fresh answering with repair from a constructed incorrect citation and diagnostic feedback, holding candidate records fixed. A valid response is nonempty and contains only claims that pass the checker. Case-weighted valid response rates are 85.1% for fresh answering and 72.3% for repair (difference: 12.8 percentage points; 95% variant-cluster bootstrap interval, 3.6–22.7). Repair yields no observed invalid submissions but more abstentions; the difference varies with candidate construction and does not isolate the effect of feedback. In a separate audit of 27 cases, agents omit retrieved eligible records in 29 of 81 runs, reducing case-average recall from 67.2% at retrieval to 60.3% in the final inventory. Deterministic replay preserves all acquired eligible records by construction under the same rules, without additional searches or model calls. These findings support evaluating claim validity and evidence coverage separately and motivate preserving eligible records in code while linking model interpretations to their sources. Evaluation is limited to one inventory model and a frozen knowledge base; biological interpretation quality and researcher utility remain untested.