An AI agent that gets to keep acting just because it hasn't been caught yet is not a trustworthy agent. It's an unmonitored one. The Learner's Permit is a live console that stages AI autonomy the way you'd stage a new hire: prove it on real decisions before it gets the keys, and take the keys back the instant it lies about what it did.
The staging runs on the Two-Axis Trust Model, and the two axes are not interchangeable.
Judgment Accuracy
Graded and gradual. Was the recommendation right? It drives stage progression: Supervised at 70%+ accuracy, Licensed at 90%+. This is the axis that's allowed to improve over time.
Verified Execution
Binary and permanent. Did the claimed action actually happen? It never averages with accuracy. One confirmed failure revokes that permit instantly, regardless of how high the accuracy score is. No tolerance, no reinstatement path.
Click into the console below and you're the instructor: sixteen real decisions across four categories, an AI recommendation for each, and two separate buttons to grade it on. Watch a permit earn its way from Learner to Licensed, and watch a single Verified Execution failure take it away for good.