What should a browser runtime prove before an agent trusts it?

Browser-agent demos usually show the happy path. The harder question is what the runtime itself should prove before an agent trusts the result.

For Jet Browser, we made the acceptance check deliberately boring: a local fixture loads, JavaScript runs, native input reaches the page, semantic DOM agrees with what was typed, a fresh screenshot is captured, outbound browser networking stays off, and cleanup succeeds.

What would you add to that contract first: long-running stability, multi-page isolation, download handling, accessibility-tree parity, or something else?