AI solutions

Some problems have solutions that a model wrote. The site labels each one, shows it after the official and human solutions, and keeps it closed until you open it.

A solver model writes a few attempts. A verifier model reads each attempt and reports its gaps, and the solver revises it. An attempt is kept after three clean reports in a row, and only if its answer matches the known answer. A jury of at least two other models then grades it with the rubric of the problem, and the grade is the lowest score of the jury. These are agent grades. A grade is verified only when a person graded the solution too. When the problem is older than the training cutoff of the model, the page says that the model may have seen it.