From past experience, experiments that perform -25% worse versus control are unlikely to recover and end up better than control when it reaches the sample size needed for significance. If the relative difference is low, i.e. less than -5% worse versus control, we wouldn’t stop showing the variant early. If the experiment affects a critical KPI such as trial starts that impacts revenue, we would turn off the bad performing variant, especially when there are 2 other variants that are performing better than control. We would review the bad performing variant to figure out why it was worse than control. In the end it’s a judgement call by the product manager taking into consideration impact to the KPIs affected by the experiment.