Layer 3 is the go or no-go decision

Green tests, lint, and review do not automatically authorize production. Layer 3 is the responsible approver’s team decision for a specific release or production action, not another test.

The model has three layers: Layer 1 checks local diff and reviewer-agent conclusions, Layer 2 runs CI checks, and Layer 3 decides release and accepts consequences.

LayerWhat it does
Layer 1Checks local diff, PR, and reviewer-agent conclusions
Layer 2Runs CI, tests, type-check, lint, and build
Layer 3Decides whether to permit release or production action

Deployment boundary

Protected branches, required checks, environment approvals, and deny rules should make bypassing the deployment boundary technically difficult. A prompt reminder is not enough.

Claude can prepareA person approves
Release notes and change listProduction deployment
Rollout and rollback planFeature-flag activation
Staging log summaryRelease announcement
PR evidence packageHigh-risk configuration changes
Operational runbookOperations on financial or critical data

Evidence package for a release gate

  • Risk classification.
  • Layer 1 review and Layer 2 CI results.
  • Staging smoke logs.
  • Rollback description and check.
  • Capability envelope.
  • Named approver.
  • Actions that are not delegated to Claude.
Traceability in a PR or release issue
Risk: review-required
CI: passed
Staging smoke: passed
Rollback: verified in test environment
Envelope: read + bounded analysis only
Approver: named human owner
Not delegated: production deploy, tag, publish

High-risk actions are not automatic

  • Production deployment.
  • Destructive database migration.
  • Secret rotation.
  • Force push to a protected branch.
  • IAM or repository-permission changes.
  • Feature rollout in a money path.

Rollback is required before deployment

For a release changing a refund flow, Claude may prepare descriptions and evidence, a reviewer agent may analyze code, and CI and staging may provide signals. The release owner still decides whether production release is permitted.

  1. Define a stable tag or another return point.
  2. Describe how to cancel the version or disable a feature flag.
  3. Assign a rollback owner.
  4. Check recovery in a safe environment.
  5. Record remaining impact on data, migrations, and integrations.
  6. Then let the approver make a separate go or no-go decision.

Canonical lesson source · JavaRush

Open JavaRush material