Auditing Update Admission in Adaptive Gene Expression Completion

We present an auditable negative boundary case for evaluating adaptive biological agents. In a fully traced retrospective replay of gene expression completion on 1,484 measured yeast deletion profiles, with a separate 156-profile stress collection for external scoring, a scripted sequential controller, with no language model or learned policy, acquires profiles, proposes refits, and admits or rejects them under a common disclosure budget. Across twenty paired partitions, campaign and fresh-reference improvement signs disagree with sealed audit signs on 13 and 21 of 120 common proposals; 2% campaign and reference gates reject 23 and 32 proposals with lower measured sealed-audit loss and do not lower mean domestic endpoint error relative to unconditional updating. Two pre-declared studies on 200 fresh partitions follow. A validation-on-demand allocator is practically equivalent to a fixed reference gate under a margin of 2% of the confirmatory endpoint; in a secondary comparison that also changes admission, the arm without reference profiles has higher external error. Among the evaluated fresh-partition endpoints, a static ridge model fitted only on the 48 initial profiles has the lowest external error, below the refined registry endpoint in 196 of 200 partitions. Fresh validation, improvement over the initial model, or improvement over zero prediction therefore does not establish a better adaptive predictor here. The evidence is conditional on one catalog, panel, threshold, batch size, and six-round horizon; acquisition is an accounting abstraction rather than an assay, and the model registry collapsed to one ridge configuration.