Question Selection and Answer Scoring Shape Revision Risk in Recurrent Language Models

An evaluation of early answers must choose which questions to release and how to read their answers. Using one scoring convention for both choices can make two procedures look similar while concealing their sensitivity to either choice. We measured this dependence on stored trajectories from Huginn and Ouro, using cloze scoring on three filtered benchmarks with 200 questions per task and cohort. The historical cohorts share questions; the three cohorts are not independent replications. At 50% release in the headline windows, changing the selection margin with the answer readout fixed changed revision risk by 9.33 to 12.00 percentage points (pp). Changing the readout with the margin fixed changed it by 9.67 to 13.00 pp. Yet changing both together gave only 0.33 to 3.00 pp. Similar risks for complete procedures therefore did not establish interchangeable selection margins. Separate sum-to-mean grids within each length unit also placed the lowest tested risks near matching rules. At the headline windows, final-depth disagreement retained the fixed-readout margin effect, although its magnitude depended on the outcome definition and observation window. Lower revision risk established neither an accuracy gain nor better stopping decisions. The results support a practical evaluation check: hold the answer rule fixed when comparing selectors, then cross the rules to measure dependence on the target.