Jewels

A jewel is an AI companion for a single method. It proposes the input a careful operator would have submitted. The method still decides: its own validation, permission, and invariant checks run unchanged, so the jewel walks through the same door as every human and can do nothing the app did not already allow. Its output is exactly the method's input, and abstaining (an honest "a human handles this") is a first-class, graded answer.

Every proposal is paired with what the human actually did and graded against it. The method's autonomy setting then decides what happens: recorded silently to learn from at shadow, queued for a person at approve, or applied outright at auto. A jewel earns its way up that ladder on its agreement rate, and each promotion is a reviewed manifest change you can revert. Because the pairs accumulate from the day a jewel goes live, they are also a training set: the Admin SDK's jewels commands turn a method's graded pairs into a small model you own, dropped into the jewel's own model slot.