Check the scoring code
These checks run the scoring code for all four tracks on fixed examples with known results. They check that the scores repeat, that changes to questions and rubrics get new IDs, and that the combined score stays consistent. The checks run in your browser without network requests. They test the code, not your AI skills.
What passing means
Each track's score() runs with checks that block the clock, randomness, network calls, and timers. The results must match saved examples. Automated build tests run the same checks. Passing means these examples produce the expected scores here. It does not prove that scores measure real-world skill or that model judges are accurate. Re-scoring uses saved judgments; asking a model to judge again may give a different result.
The checks have limits. They block common global functions, but are not a sandbox. They can miss references saved before a check or code imported when a module loads. They also check only this JavaScript runtime. Identical results across engine versions are not yet proven, and score records do not store the runtime version. See how scores can be checked later.