Monitoring
Overdue and no-data alarms
The two failures nobody notices from inside the results screen: a test still running past its end, and a test that quietly stopped collecting.
An overdue experiment was scheduled to finish and did not. It is still running, still splitting traffic, and still costing you the audience it takes, usually because the decision it produced was never made. It is ranked as a decision waiting rather than as a fault, because nothing is broken and somebody just needs to call it.
A no-data experiment is the more urgent one. Nothing recorded for a full day while the test is still running almost always means a targeting rule stopped matching, or the snippet went missing in a deploy. From inside the results screen this looks like a quiet week; from outside it is a test that has stopped being a test.
Both reach you rather than waiting to be found. Both are on by default for email and Slack, along with the split mismatch, because health signals earn an interruption in a way that lifecycle events do not.
A no-data reading also blocks automated stopping. Acting on a conclusion drawn from an experiment that stopped collecting is exactly the case the trust gates exist for.