Each instance is a closed-loop sequence of QPs sharing one model. Its time is setup (model parsing and solver setup) plus every step, measured by the supervisor that reveals one parameter vector per step; the score is the shifted geometric mean of instance totals. A trajectory passes only if every step meets the primal, objective and applied-solution tolerances. Step statistics (median, p99, max) and setup are shown for context; they are not scores.

Shifted geometric mean = exp(mean(log(time + shift))) − shift. Recorded failures contribute the task’s time limit. Missing measurements make the score incomplete. Dev and heldout are scored separately; All combines every instance in both splits.

Per-instance best scores are best-observed composites and may combine revisions. Selected revision uses the recorded development selection and its matching heldout evaluation. All is a derived summary, not an official benchmark selection score. Comparisons are grouped by recorded dataset identity and task settings; inspect run configuration for hardware differences.