Build a conformed table blind from intent + raw source
The AI writes its dbt SQL from the dashboard intent and the raw landings, and self-heals against answer-key-free checks — it never sees any existing dbt definition or layer output. The GO/HOLD verdict comes from those gates: grain, source ties, data quality, and intent coverage. Migrating from an existing layer? Score any finished build cell-for-cell on Verify vs existing layer.
Copies only the raw tables this output needs from the client source into our warehouse. Everything after this — profiling, the AI's questions, validation and the dbt build — reads the landed copy, so the client's system is touched once and read-only.
No dashboard intent yet — add Sisense widget exports (*.jaql.json) or a markdown mockup (*.md) of the desired output so the AI knows what the dashboards need.
For each table: the AI asks anything it would otherwise guess (only if there are open questions), writes the SQL blind, self-heals, and scores a GO/HOLD verdict from the answer-key-free gates.