Teaching evidence only. The experiment outcomes were deliberately manufactured. They do not establish a real treatment effect. Public registration and payments remain disabled.
01 / A defined funnel
Unique eligible users · First-start attempt · Strict order · Inclusive 72-hour deadline
Start: 12 usersLesson: 8Exercise: 5Overall: 5/12 = 41.67%
Conditional: 8/12 = 66.67%; 5/8 = 62.50%.
Recorded stopping points do not explain motivation. Repeated or out-of-order actions do not create extra converters.
02 / Source reconciliation
Event delivery records, not people| Raw rows | 45 |
|---|
| Exact delivery retry | −1 |
|---|
| Internal event | −1 |
|---|
| At/after watermark | −2 |
|---|
| Accepted | 41 |
|---|
13 signup records − 1 internal account = 12 eligible users. Source assumed complete strictly before 26 January 2026, 00:00 UTC.
03 / Compare the same product age
Mature day-7 return: 3/8 = 37.50%. Day 7 means [signup + 7 days, signup + 8 days). The latest cohort is unobservable, not 0%. The whole cohort row is withheld until every member has complete follow-up.
Small observational cohorts. No causal or forecast claim. Day-8 returns do not count for exact day 7.
04 / A positive estimate with substantial uncertainty
Difference: +8 percentage points · Relative lift: +40%.
Approximate 95% interval: −3.79 to +19.79 percentage points.
Independent-user, fixed-horizon unpooled normal approximation. Sparse outcomes are refused by the reference; no cluster, sequential or equivalence inference is supplied.
100/100 passes the illustrative binomial SRM check. That diagnostic assumes independent 50:50 assignment; this fixture uses a fixed-count balanced shuffle, so a real check must match its allocation design. Illustrative sample plan: baseline 20%, absolute MDE 8 points, two-sided alpha .05, power .80 → approximately 447 per arm. Error counts are too sparse for this kit’s normal interval; they do not establish safety.
05 / What to do with this evidence
Investigate recorded first-step progression and event delivery. Keep the retention denominator restricted to mature cohorts. Do not call the manufactured treatment a winner. A real follow-up needs a reviewed assignment design, adequate sample, complete observation and a predeclared guardrail decision rule.
Read all eight lessons · Compare worked solutions · Read the worked decision memo