Experiments

How experiments are judged

The current version always runs alongside the challengers, so results are compared fairly.
A test is called once, at its planned end: a version wins only at 95% confidence or more, with at least 20 conversions in every version.
Before the planned end, a lead can still reverse, however strong it looks, so the test isn't called early.
If nothing clears the bar at the planned end, the current version stays.

Running now · 0

None running. The agent proposes new ones in its stories.

None yet.