We're investigating reports that some cohorts are calculating incorrectly. This has a knock-on effect for feature flags that depend on the failing cohorts, causing them to evaluate false.