Technically it’s how unlikely of an event under a null model.

The null model is key: if you are mischievous, you can just define a seemingly benign but incorrect null model and generate extreme p values without discarding: all values will be significant!