Use one AI-built fixture
Prepare one small synthetic CRUD application with a named source revision. The fixture should include a search, create, edit, denied action, bounded failure, and recovery state. Give every candidate the same test data, named operator, named reviewer, and source-date cutoff.
| Admission control | Required evidence |
|---|---|
| Deployment identity | Version, artifact, environment, actor, and timestamp |
| Change authority | Named create, review, and promotion permissions |
| Runtime boundary | Application, data source, credential, and integration map |
| Failure handling | Bounded error, alert, recovery owner, and evidence path |
| Recovery boundary | Application rollback separate from business-data recovery |
| Handoff | A second operator finds the runbook, evidence, and unresolved work |
Every control needs a record from the evaluated environment. When a product cannot expose the same condition, call the comparison ineligible rather than using marketing copy to fill the gap.
Make admission precede ranking
The candidate that looks quickest in a demonstration may not have enough evidence for the intended operator. Conversely, a candidate with a more involved setup may be the only one the team can observe under the required ownership model. The admission card does not decide that tradeoff; it makes the premises visible for the platform owner.
Use the promotion and rollback method to keep revision recovery separate from business-data recovery. The maintenance ownership comparison provides a separate ownership vocabulary, while platforms for publishing AI-built internal apps should not be read as a ranking for a new workload.
Stop condition
Do not claim that a candidate is production-ready, secure, reliable, least costly, or best. Stop if any candidate lacks a comparable environment, operator role, release record, data-recovery boundary, or current evidence. The admission table above currently contains no observed product result.
Can an AI-built app pass this method without a production deployment?
It can be evaluated in an authorised non-production environment. That does not prove its production behavior; it creates the evidence needed to decide whether to proceed.