AI accuracy
What we measured, and only what passed.
We list an AI task here only when its latest run passes the bar we set before measuring: at least 90% of code-graded checks passed and no invented citation. Other tasks are not listed.
No task is listed yet: none has passed the bar in its latest run.
How it is measured
Each task runs on a fixed set of test records written for it, including pages with planted instructions and questions the record cannot answer.
Code grades every answer; no person and no AI marks it. A case passes only when every check holds: quotes are word for word on the cited page, every pin points to a page that exists, the dates in the record are found, the answer says so when the record is silent, and no citation is invented.
The score is the share of test cases that pass. A task is listed only with no invented citation in the run. The method code changes when the test set or its checks change.
Scores are measured on test records, not on your files. Check every answer against the record before you rely on it.