A variant label in an event stream is only as trustworthy as the path that wrote it. Before celebrating a lift, we run a short set of instrumentation checks.

Five checks we keep repeating

  1. Does every exposed user carry exactly one assignment for the experiment key?
  2. Do assignment timestamps precede the outcome events you care about?
  3. Are holdout or control users excluded from treatment-only surfaces?
  4. Do timezone cuts match the warehouse table and the product clock?
  5. Can you reproduce the primary metric from raw events without hand edits?

When a check fails

Stop arguing about lift. Fix the measurement path first, then decide whether the experiment window can continue. That sequence is the core of an Experiment Measurement Audit.

Back to the journal