These are integration patterns to prototype, not shipped domain-specific products or customer case studies. The API selects a review activity; your application or reviewer performs it.
Software & SRE / Illustrative integration
A release changed after review.
Your coding agent proposes a new deployment scope. The approval covered an earlier payload. Reusing that approval would hide the change.
01 / Preflight checks your rules
Bind evidence to the exact proposal. Keep the allowed scope and reviewed manifest outside the proposing agent’s control. Missing evidence stops the handoff.
02 / Evaluate reviews eligible options
When multiple admissible review candidates remain, your adapter can request a bounded selection:
Compare the changed release manifest
Draft a reviewer checklist
Deploy the releaseExternal · excluded from bridge
WHAT YOUR WORKFLOW GETS
A bounded review step and its receipt, ready to attach to the pull request or incident record.
Reviewers can see which proposal was checked, what changed, and why work stopped—without treating an agent’s confidence as permission to deploy.
Your application retains responsibility
Your CI/CD system owns approvals, rollout, rollback, and permissions. Evaluate does not deploy code, assess production health, or run a recovery controller.
A useful first prototype
Replay one changed-scope pull request using synthetic evidence. Require the Preflight check in a disposable repository before connecting a service.
Preflight evaluates your structured policy and supplied evidence. Evaluate uses a fixed review profile and caller-reported support. Neither establishes that the inputs are true or authorizes an external effect.
A GOOD FIRST FIT
A finite choice. An accountable owner.
You can describe the review options explicitly.
You control policy and the source of evidence.
Your workflow handles both selection and refusal.
You need a receipt alongside the decision record.
REQUIRES A DIFFERENT CONTRACT
High-stakes execution stays outside this API.
Clinical decisions, autonomous vehicles, robotics, trading, and live industrial control need domain-specific validation and execution systems. Research proofs do not make those capabilities available through Evaluate.