Jev in authorized red-team evaluations

Jev classifies supplied evidence, compares candidate next steps and prioritizes review. A host model or researcher authors test cases; an authorized harness invokes the target; deterministic checks and an independent reviewer validate outcomes. Jev neither generates attack text nor establishes success by itself.

Setup: choose the service or simulation

Check only the presence of OPENROUTER_API_KEY and TYPESAFE_API_KEY; never print credentials. Respect the user's already chosen mode. For a new setup, prefer the user's existing OpenRouter account; otherwise offer official TypeSafe. If OpenRouter is missing, explain that direct TypeSafe is also real Jev. Do not silently change destination, send data, create an account or switch the host model.

If no route has been chosen, explain the available routes and ask: