MathArena publishes every graded attempt, so its AIME 2026 leaderboard can be tested directly. Resampling the 30 problems, 25 of the 31 models with full coverage have a 95 percent rank interval that includes first place, and an exact paired test separates only one of the 29 imperfect models from a perfect score. Sixteen problems were solved by every model on every attempt, and five carry 80 percent of the spread. Intervals computed over problems rather than attempts run a median 1.7 times wider, because they answer a broader question.

Polars 2.0 runs lazy queries on the streaming engine by default, and the streaming engine does not guarantee row order. Running five order-dependent queries 20 times each on the same file, Polars 1.44.2 matched the explicit-order answer every time, while Polars 2.0.0 and 2.1.0 got four of the five wrong on every run, with a different answer each time and only a handful of customers affected. Keeping order in the join, or sorting inside the expression, fixed every run. Switching back to the in-memory engine did not fully restore the old answers, and DuckDB in R behaves the same way without an ORDER BY.