Explainable AI model

Explainable AI narrates gate state and evidence in plain language. It runs against an inference endpoint the administrator selects. It is never in the decision path: no verdict, no decision, and no authorization depends on it.

Selecting a model

Open Control from the profile menu. The model panel lists every model the inference endpoint offers, grouped by source: self hosted models and cloud models. Choose one to make it active. The active model is used for every narration in the console from that moment.

Testing a model

The panel includes a prompt field. Enter a short question, run it against the selected model, and read the answer and the response time before making it active. A model that cannot be reached shows the endpoint that failed.

What the narration is grounded in

Every narration is built from the scan record for the run: the gate, its evidence, the scanner output, and the developer's justification. The prompt carries the evidence; the model is asked to explain it, not to add to it. A narration that cannot cite its evidence is not shown.

What it never does

It never proposes a basis for accepting a risk. It never records a decision. It never signs. Nothing it writes is part of the authorization record.

Endpoint

The console reaches the inference endpoint through the service's own proxy, so the browser never holds an inference credential. In a Program deployment, the endpoint is inside your boundary.