Every prediction is logged when it is made and scored once the class is seen
calling dogs. Too late means the class started before the time
promised — the dangerous kind, because anyone trusting it would have
missed their run. Too early means it was pessimistic and you
would have been stood at the ring waiting.
Each class counts once per row, as the average of its own predictions.
Averaging over raw predictions instead lets a class that sat in the queue for an
hour — and was polled forty times — outvote one predicted twice, which is how a
single ring having a bad afternoon came to look like a systematic failure.
No scored predictions yet. Predictions can only be scored once the class they
refer to has actually started, so this fills in over the course of a show day.