Choose checks that can tell whether the work meets its requirements. Unit tests, contract tests, visual comparisons, benchmarks, and formal verification each answer different questions. Combine them with review where meaning, usability, or a consequential decision needs a person. Keep the claim, the check, and the artifact together so confidence does not depend on the agent's final message.
What this makes possible
- Machine-checkable acceptance criteria and reusable test harnesses
- Positive examples, negative examples, and failure reproduction
- Review workflows connected to the relevant change
- Evidence records that identify what was checked and when
On the production floor
A compatibility factory checks accepted inputs and rejected inputs against the same contract, then presents any meaningful behavioral differences for review.
How you know it works
Challenge the verifier itself with known bad outputs. Re-run the relevant checks when their code, configuration, or acceptance criteria change.