Choosing a genetic code-repair configuration requires separating favorable average behavior from exposure to severe losses on particular tasks. This article asks how configuration choice changes when fitness and parent selection interact and the available evidence consists of rounded median test fractions. The analysis constructs 172 configuration records covering 43 eligible task–model pairs, retains all four combinations of tournament or lexicase selection with count or count-and-discrepancy fitness, and calculates factorial contrasts, precision intervals, and minimax regret. Seven contrasts are positive, three negative, and 33 zero at the printed resolution; six positive and three negative signs remain resolved throughout the precision intervals. The largest opposing contrasts are 0.94 for Paired Digits with CodeLLaMA 7B and -0.50 for Solve Boolean with GPT-4. Lexicase with count-and-discrepancy fitness has the highest mean fraction, 0.26535, but maximum regret of 0.50. A randomized choice over configuration-level median scores reduces that maximum to 0.23967 while lowering the mean to 0.25354. Accounting for display precision gives a worst regret of 0.24586. Task-deletion evaluation increases the nominal randomized rule’s maximum regret to 0.36167. These findings establish a finite-evidence trade-off between average score and protection against observed configuration failures. They do not establish the performance distribution of randomized repair executions or guarantees for unseen software.
