VERIFYDevtools & code
Jev by TypeSafe AI
o8@o8dotrun
1:33Jev is a referee mechanism by TypeSafe AI that records answers as receipts and replays them against previous events before allowing them to gate anything.
One setting, off by default. When it's on, a typed judgment model answers yes-or-no questions about the diff I'm looking at. It never writes anything and it never decides anything.
Here it reads a worker's change before the operator does: docs only 1%, adds tests 2%, placeholder data 88%, risk 1.1 of 4. The card says what that is — advisory, not used by the merge decision. The merge button is still yours.
Lead is Claude Opus 5, the worker is DeepSeek V4.1 Flash through OpenCode 2, the referee is Jev by TypeSafe AI.
Every answer is recorded as a receipt, and we replay those receipts against what happened before any of them is allowed to gate anything. Right now none are.