The analyst's mental model doesn't exist anywhere official.
The trusted set lives in people, not in documentation.
Every experienced analyst has a mental model of the warehouse that does not exist in any official document. They know which tables are canonical, which are deprecated, which are context-dependent, and which will silently mislead you. This knowledge is the real interface to the data.
mrr_monthly: source of truth, Finance-approved, use for anything board-facing. events_processed: deduplicated, QA-checked, 15-minute lag, safe for product metrics. active_users: correct definition of active (not users, not user_events). These distinctions exist in the analyst's head, not in the catalog.
raw_events_ingest: avoid. Duplicate events from SDK retries, bot traffic included, never use for metrics. events_v1: deprecated, frozen at migration, data stops 14 months ago, still queryable. user_signups: misleading name, contains all historical signups including cancelled accounts.
New analysts learn this map through osmosis, through incidents, through questions asked in Slack and answered by whoever has been there long enough to know. The map is accurate and maintained. Any automated system that hasn't been through the same apprenticeship has no access to it.
The analyst's mental model is an operational asset. The question isn't how to write it down more thoroughly. It's how to make it available to the systems that now need it.
