Interaction detection

This is the failure mode that is almost impossible to diagnose from inside the experiment reporting it. A redirect experiment pulling traffic out of another audience surfaces as a split mismatch in the other experiment, where nothing in the data points back at the cause. Somebody spends an afternoon investigating a test that is working correctly.

Interaction detection reports the relationship rather than the symptom. It names both experiments and what they are colliding over, so the alarm arrives with its explanation attached rather than as a mystery in the wrong place.

It also catches the case that produces no alarm at all. Two experiments editing the same element keep a perfect traffic split, so the sample-ratio check sees nothing wrong, and what visitors get is a page combination neither test intended. That is why interference is a separate trust gate from the split check rather than being folded into it.

The finding blocks automated action on the affected experiment for as long as it stands. The reason is worded once and reused, so the board editor, the engine and the attention queue all say the same sentence about the same pair rather than three phrasings of it.