Learning What to Measure Next: Failure-Aware Closed-Loop Evidence Allocation for Protein Reliability

Protein design increasingly produces more computationally plausible candidates than can be physically measured. We formulate reliability screening as a sequential evidence-allocation problem and instantiate it as a traceable closed-loop decision agent with explicit state, action, measurement, observation, update, and evidence-registry semantics. Using folding stability as a controlled endpoint, the original frozen acquisition policy combined predicted failure relevance with bootstrap-ensemble disagreement. A follow-up component ablation isolates the two terms. On a fixed 30,000-sequence MGnify Stability replay, Failure-Only selection based only on predicted low stability reaches 69.50 $\pm$ 0.47% low-stability recall at 8,192 measurements, compared with 53.95 $\pm$ 0.77% for the original joint PRIB score, 27.62 $\pm$ 0.81% for Random, and 27.55 $\pm$ 1.01% for generic Uncertainty. Failure-Only exceeds the joint score in 24/25 seed-by-budget cells. With five ensemble members, the ordering remains stable and Failure-Only reaches 69.87 $\pm$ 0.99% versus 53.51 $\pm$ 0.50% for the joint score. A same-state counterfactual diagnostic localizes the effect to acquisition-score choice: five-member Failure-Only improves immediate failure yield over the joint rule by 14.21 percentage points on average across the 25 seed-by-state cells. Across the three active scores the ordering tracks objective alignment: failure relevance alone is strongest, the equal-rank combination is intermediate, and disagreement alone does not outperform Random at any tested budget. These results support objective alignment as a first-order design choice for closed-loop scientific agents that allocate scarce physical evidence.