HYPOTHETICAL ONLY — redesigned onboarding versus original.
Eligible unique new learning users, exclude preidentified internal accounts.
Randomize stable user once50:50, persist original arm; inspect duplicates/crossovers.
Primary: any completed practice in[assignment,assignment+7 days),binary per user.
Analysis: original assigned users, including non-exposed; complete observation required.
Missing/unknown telemetry stops interpretation, not silently assigned failure.
Guardrail: any error in same window. Real plan must set an acceptable harm margin
and a separately validated sparse-data inference method BEFORE interpreting results.
Illustrative planning:p0=.20,MDE=.08 absolute,alpha=.05 two-sided,power=.80,n≈447/arm.
Fixed enrollment target plus full7-day follow-up; real calendar end depends on traffic.
No optional stop when significance appears. Safety stops require predeclared rules.
SRM diagnostic exact50:50 binomial,p<.001 alerts. It cannot validate every assignment defect.
One primary outcome; exploratory segments labelled exploratory, multiplicity handled separately.
Decision: no automatic launch. Review validity, uncertainty, practical benefit and harm.
Do not claim this incomplete hypothetical guardrail plan authorizes a real launch.
Fixture only 100/arm and deliberately manufactured20/28 outcomes; inconclusive.
No actual visitor randomization, external analytics service or payment feature enabled.
