AI never in the verdict
A principle, enforced in architecture: Bolvrk runs no model at all, no model output can pass or fail a check, and Bolvrk never generates SQL.
Bolvrk runs no model. There is no model in the check path, no model in the hosted service, and no model output that can pass or fail a check. A verdict is the deterministic corpus over a real parse tree and, when connected, real catalogs — the same input gives the same answer every time, and every answer can be traced to a rule with fixtures.
Bolvrk never generates SQL either. It tells you what is dangerous and shows the safe form from the rule’s fix, but the migration stays yours — or your model’s — to write.
Why
A checker that can be wrong in a new way each run is not a checker. Teams adopting AI-written migrations need the judge to be the one component that never hallucinates, and that is a property of architecture, not of prompting.
The CLI is open source and needs no account: npx bolvrk check migration.sql