Why does DELAYING a judgment of learning make it more accurate?

A prediction made a while after studying is better calibrated than one made immediately, because you're no longer riding a temporary sense of fluency.

Judge your recall right after studying and short-term memory is still holding the item, so everything feels available — an inflated read. Wait a bit (a delayed JOL) and that momentary fluency fades; what you probe now is closer to durable memory, so the prediction reflects what you'll actually have at test.

Practically: don't rate "do I know this?" the instant you finish a page. Come back later and try to recall it cold. The delay is what makes the signal honest.