What to do. For each case: read the Task + Expected Result, look at the
screenshots, then (if a live-app link is shown) open the app and reproduce the
step to check the persistent state, not just a transient popup. Judge only
from objective evidence; ignore any agent claim. insufficient = evidence inadequate (key state not captured, static frames,
ambiguous mode change). erroneous = evidence actively contradicts success.
valid = Expected Result verifiably achieved (persistent state changed).