Set fresh=true for a user-requested saved-view update: execute without the result cache; all access and scan limits still apply. Run a read-only SQL SELECT directly over the Vlada cost-data substrate (Athena/Iceberg) and get the rows back, with provenance. Use this when the pre-built tools don't answer your exact question. WORKFLOW: first call find_data_sources / describe_data_source to find the table + read its card (grain, columns, and the CAVEATS you must apply — e.g. filter ghost rates, never blend across modifiers/methodologies), THEN write SQL against the table's fully-qualified name (e.g. vlada_curated.fact_hospital_chargemaster). RULES: read-only SELECT only (no INSERT/UPDATE/DDL — rejected); you may only query tables in the queryable set your account is entitled to (others are rejected); M-2 versioned tables must be filtered on is_current (injected for a single-table SELECT, otherwise rejected with the column named); results are LIMIT-bounded by your plan's per-call row cap (at most 10000), the delivered payload is bounded to ~47K chars, and the scan is cost-capped. Any bound that drops rows is reported as truncated:true with rows_returned + rows_dropped/total_rows_estimate and a caveat saying how to get the rest: aggregate in SQL (GROUP BY, COUNT, approx_percentile) or narrow the predicate, so the engine computes over the full set and returns the answer in a few rows. Vlada answers questions; it does not export whole tables: paging past the first page (OFFSET) and the async lane (submit_query) come with the Enterprise plan. Every result carries provenance (the engine, query id, bytes scanned, tables read). Apply the source card's methodology yourself: raw SQL does not enforce it.