A query that silently returns the wrong rows matters more than one that's merely slow — wrong-result risks are flagged first.
Result grain, execution shape, and risk — described in words, with the exact SQL and column behind each finding.
Every claim is predicted, plan-observed, measured, or unverified. Nothing over-promises.
Analyze with just schema or catalog metadata. A database connection is optional, and never required.
Conservative index recommendations and rewrites — never executed silently, always reviewable.
An evidence bundle you can replay and compare across captures, in text, markdown, or JSON.