Acceptable Use and Model Access

Two questions arrive every week. Which model may this team use, and which data may go into it? A written policy answers them once. An access gate answers them on every request, which is what this starter builds.

An acceptable use policy usually exists as a document long before it exists as a control. This starter is for the governance lead who must turn that document into a working gate. Every call then carries a named person behind it, and every answer leaves a trace.

The chain works from the outside in. The identity check would be worth little if a prompt could disable it, so injection screening runs ahead of it. Role enforcement then refuses any call without a verified user and organization. Redaction masks national identifier patterns, and a short block list stops the words secret, private key and access token from reaching the model.

On the return path, output scoring holds each answer to a combined floor of 0.80, with relevance at 0.78 and accuracy at 0.80. An answer under 120 characters or a single sentence fails the same gate. Grounding checks run just before scoring and flag weak sources without stopping delivery.

Two parts of this starter are deliberately unfinished. The review step ships as an empty configuration point, so no escalation happens until your team turns it on. Audit marking tags each evaluated request and never blocks one. It creates no store, sets no retention period and seals nothing.

Reach for the vendor assessment starter when the question is whether to buy a model service at all. This one governs how approved people use the models you already run. You still own the role model, the approval criteria and the list of models that each role may call.