01 The ticket
Edit it freely — the state is prefilled once, whatever it says.
02 The questions
Write your own. Every criterion is asked about the same state, which is prefilled once — so a criterion costs a short suffix, not another full prompt.
03 Run
Fetching weights. First visit downloads once, then it is cached.
04 Elapsed
Lanes are measured separately and aligned at zero. Running both at once on one GPU would make both numbers meaningless.
Readout
probabilities taken from the logits
—
Generation
same model, same question, writing JSON
—
05 Readout
One softmax per criterion, over that criterion's own options.
Probabilities are conditional on the options shown and are not calibrated confidence. Don't threshold on them.
06 Generation
The same model, shown a worked example, asked for JSON.