Keep collecting data until you can model precisely the incremental data you're collecting.

The surprisal is gone when the model is right.

This only works if you're actually getting disconfirming evidence.

You can get the disconfirming evidence from a truly random sample – both confirming and disconfirming evidence.

Another approach is to narrow in on just the disconfirming evidence to update your model faster.

But be careful: disconfirming evidence hurts, and so if you're applying a selection pressure to what evidence you actually receive and act on, you're almost certainly filtering out disconfirming evidence.

This is especially true in a high-kayfabe environment.

Especially an environment with high individual downside and a top-down plan.

If you get evidence that shows the top-down plan is wrong, that will cause a lot of pain for you!