Growth Analyst

1. 1 Frame the question Claude Opus 4.8 1090ms

   "Improving" is ambiguous. Retention by signup month conflates a product change with a channel change — reading by period instead.

2. 2 Build the table Claude Haiku 4.5 610ms

   9 cohorts × 12 periods. Northwind runs a 3-month horizon, so truncating there rather than extending to 12.

3. 3 Audit for bias GPT-5 870ms

   The Sep cohort has 22 members and the best curve on the chart. That is noise, not a win — showing it, but not concluding from it.

Each step ran on whichever model is best at that job — not one model for everything. The worker is the workflow.

• Read: warehouse tables and the analytics standards doc — through a read-only role, one schema at a time
• Write: analysis workbooks and a documented query log only; source tables are never written to
• Escalate: the analyst when a cohort falls under 30 members, or when a metric’s definition is ambiguous
• Never: mutate, delete or backfill a source table under any instruction
• Never: draw a conclusion from a thin or survivorship-biased cohort without flagging the caveat

• Week-4 retention up 4.1pp — real, and visible across 7 of 9 cohorts.
• Sep cohort (n=22) excluded from the conclusion. It was the best-looking one.
• Read by period, not signup month — the framing the question actually needed.

private rules were taught in one workspace and never leave it. general rules are true about the work, so every copy of this worker gets them.