Collision detection

Concurrency costs nothing in sample size here, so the platform does not limit how many experiments run at once. What it does instead is check the narrow set of ways two experiments can genuinely damage each other, and report exactly three things.

A selector collision means two variants mutate the same element, producing a page combination nobody built or reviewed. A redirect collision means traffic moves into or out of another experiment’s audience, breaking the independence both readings rely on. Both block. A shared goal is reported as a warning rather than a block: the effects may not be additive, so neither result is guaranteed to replicate alone.

Disjointness is checked before any of that. If no visitor could be in both experiments, because they target different devices or provably different URLs, nothing else is reachable and the pair is safe whatever it touches.

One honest limitation, worded carefully in the product rather than glossed over: selector matching is exact after normalising whitespace, which makes it a detector of likely collisions and not a proof of safety. Two different selectors can address the same element and will not match. A hit is strong evidence; a miss is not evidence of absence, and nothing in the console calls a pair safe.