Accuracy and guarantees
How checking works, what we measured, and where the limits are.
How checking works
Independent answers, cross-checks, then an editor.
- 1Models from different companies answer on their own.
- 2Their answers are split into statements and similar statements are grouped.
- 3Other models check each statement against your sources and must quote them.
- 4Disputed points get further rounds; an editor writes the answer from what held up.
What we measured
A preregistered test on 483 unseen questions.
On short, checkable questions Smart mode matched the best single model (Claude Opus 5.5) and was significantly better than GPT-6.1 Sol, at under half the cost of the Standard council. The full results are on the Accuracy page.
Note: On long free-text answers the test did not show an advantage over the most careful single model. We say so on the Accuracy page and do not claim it.
Calibration and Confirmed
Turning checks into a measured error rate.
A calibration labels a sample of a workspace's statements by hand and sets thresholds so the error rate of Confirmed statements is bounded. It is signed off per workspace, depth and model set (Pro Max).