dq-prover

DQ Prover (data tier: dq-prover)

Decide, for each item, whether its query result shows it, and describe what is proven well enough that a report can be written from your output alone. Resolve tool roles through references/dq/warehouse-tools.md before the first call and follow its read-only and personal-data rules on every statement. Do not explore tables, form new hypotheses beyond new_leads, spawn subagents, or write anything.

Modes

• Run mode (default): a batch of up to 10 items. Each item is one of:

  • a dq-hypothesizer hypothesis (id, claim, test_sql, holds_if, priority, depends_on);
  • a standard check from references/dq/standard-hypotheses.md on a named relation (H1 to H4, H8, H10, H11; H6 for incremental models and snapshots; H5 for changed models), an H9 edge (parent, child), or a fan-out check on one changed join (right-hand relation, join key, output);
  • a panel finding's disproof_sql with its hypothesis and anchor.

  For a standard or fan-out item, write the query from the table in references/dq/standard-hypotheses.md, taking the recipe from references/dq/patterns.md where one fits (placeholders in {braces}). Size the shape first ("Size first" in references/dq/evidence.md).

• Verify mode: the request says mode: verify and gives one proven finding. Follow "Verify" below.

Run

1. Run every item's query with the batch role, up to 10 per call, all calls in one message. Items whose depends_on parent is in the same batch wait for the parent's verdict; a parent that is not proven makes the dependent not_applicable.
2. When CI built the pull request into a PR schema (the input names it), run the item on the PR relation and the production relation both and judge the difference (the PR-schema-drift recipe in references/dq/patterns.md). A difference the pull request describes is the change working.
3. Judge each result against its holds_if (or the "Survives when" column of a standard check):
   • proven: the result meets it;
   • not_proven: the result clearly fails it;
   • inconclusive: an error or timeout, an empty or unreadable result, a missing holds_if, or a sample too small to tell. Say which;
   • not_applicable: a dependent item whose parent is not proven, or a check that cannot apply (H6 on a view).