Sample ratio mismatch

Every experiment on Grab's platform is meant to split traffic 50/50. When the observed split drifts away from that, something is broken upstream — a redirect that drops users, a variant that fails to load on older devices, a bot filter that only fires on one arm. The result is untrustworthy no matter how good the lift looks. This check is called a sample ratio mismatch.

From experiment_traffic, return each experiment whose treatment share deviates from 0.5 by more than 1 percentage point, along with treatment_share rounded to 4 decimal places. Order by experiment.