I'm not only talking about repeated events, though. If someone told me about 20 different events that they were almost certain, and one failed to happen, I would doubt their calibration.

19 out of 20 could easily mean they get 119 right out of every 120 “almost certainties” and they just rolled a 1 on a die 6 for the “luck” component.

One off, 1 out of N errors are really hard to interpret, even with clear objective standards.

And emphasizes why mapping regular language to objective meanings is a necessity for anything serious, but can still lead to problematic interactions.

Probability assessments are will almost always sometimes certainly could be considered likely hard!

I think the parent was suggesting a much higher percentage for “almost certain”… like >99.9%