Quality gates
Quality is enforced by evidence, not by asking the agent to "be careful".
The test ladder
| Layer | What it proves |
|---|---|
| Unit | Logic and edge cases in isolation |
| Integration | Boundaries: database, queue, object storage |
| End-to-end | The user flow works through the real surfaces |
| Visual | The UI matches the approved design baseline |
| Security | No new secrets, vulnerable dependencies or forbidden changes |
The project's test strategy defines which layers are required. The e2e command runs in the sandbox after unit tests; a failure blocks the PR.
Acceptance evidence
Every task has acceptance criteria. A development run must produce an AIP_PROOF artifact that maps evidence to each criterion:
criterion: user can sign in with SSO
evidence: e2e/sso.spec.ts > signs in with SAML
The platform validates that the referenced test or file exists in the diff. A proof that points to a missing test fails the run.
Design checks
For UI changes the platform can compare a rendered screen against the project's baseline:
- layout, spacing and typography
- component usage from the design system
- accessibility checks
- visual diff threshold
A visual difference does not approve or reject itself. It produces evidence for the human reviewer.
Diff policy
Project rules can block changes before review:
- forbidden paths (secrets, generated files, infrastructure)
- maximum diff size
- dependency changes that need an ADR
- missing migration for a schema change
- new network calls in the sandbox
Review findings
A review run reports findings with severity. Findings are not automatically blocking unless the project says so; the PR approval remains the human gate.
What to do when a gate fails
- Read the failing check, not just the summary.
- Reproduce locally with the same command.
- Fix the cause, not the assertion.
- Re-run the check or the agent run.
- Record an override only when a human accepts the risk; the override is audited.
Next
The clean audit
The clean audit re-runs the project's rules and the architecture checks over the merged code, independent of any run. Run it from the project's quality page. A clean audit is the proof that the rules still hold after a burst of merges; the result is recorded in the audit log and never rewritten.