The question.
Code can look plausible without handling the cases that matter. Comparing AI systems requires a way to examine correctness and understand how the result was measured.
Our approach.
Deadline tests Python, JavaScript, and SQLite against hidden cases. Correctness is the headline measure; token efficiency is separate. Versioned results keep changes in grading visible.
What you can do.
Explore official results, compare effort settings, and read the methodology behind each score. The current official results use one attempt per task, so small differences are not enough to establish which model is better.
Read the methodology