OPS.16:5.3 - An AI-assisted software operation stops at a provider boundary

A software operations group has independently admitted AI-ReleasePrep-v3 : U.Method for one release-preparation family. Exact AI-ReleasePrep-Description-e4 : U.MethodDescription states the human review, evidence capture, security check, fallback, and release-authority boundary. A prompt repository, model endpoint, and usage counter are supporting Systems or records; none is the Method or proof of enactment.

The trial WorkPlan compares the current human preparation route with admitted v3 for a small set of eligible low-risk changes. It names actual human reviewers, the admitted Method expected to be enacted, model and provider condition, security and software-assurance checks, fallback, review burden, evidence window, and a stop if the provider or model condition changes. Actual release-preparation and review Work is admitted separately. Model calls, traces, generated text, test results, security observations, human decisions, and release authorization retain their own subjects.

For eligible cases under the qualified provider and model condition, the observations can support a bounded adopt or branch decision if human review, assurance, service, and burden conditions remain current. If the provider changes the model to an unqualified edition during the interval, the existing evidence no longer supports that branch. The team stops AI-assisted starts for the changed condition and uses the named fallback while qualification is unresolved.

The return preserves the admitted Method and its exact description while narrowing its applicable branch. It establishes neither model capability in general, causal superiority, release authority for the tool, nor transfer to high-risk changes. Security, software assurance, model capability, employment, and release decisions remain with their owners.