OpenWork Jev Verification
OpenWork uses Jev as a typed verification judge for agent-produced work in its evaluation testkit.
OpenWork uses Jev as a typed verification judge for agent-produced work in its evaluation testkit.
OpenWork’s testkit can use Jev as an evaluator when compiling a verification intent against a dictionary of checked-in checks. The evaluator passes the compiler’s state and questions to the configured gateway; the selected checks are then replayed through real deterministic steps without further model calls.
Does the user’s intent require the composer check? Judge which checks were requested, not whether the app already passes them.
A draft-review user intent and a versioned dictionary of requested checks, including a check that the composer is editable and says “Review notes”. The app screen itself is not sent.
Provide uniquely identified checks with author-written descriptions and executable observation bindings.
Ask Jev whether the dictionary covers the intent and which checks are relevant or necessary.
Reject incomplete or malformed coverage and inspect the selected fixed checks.
Execute selected checks in dictionary order against current application state without calling Jev again.
Jev selects from a closed set of author-defined checks and cannot create assertions or perform the verification itself; the resulting plan remains reviewable and replayable.