Deal a round face up and call the cards as you normally would. Each
zone shows when it first produced a value and when it last changed;
a zone counts as settled once it has held the same value for
3s. Time to settle is the number
that matters — a fast first guess that flips later cost the
table the later time, not the earlier one.
Test Recognition
Green = local model answered, with its
confidence. Yellow "Claude" = the local
model came in under the bar and Claude answered instead — Claude
returns no confidence of its own, so the figure shown is the tokens
that call cost. A /n suffix means one batched call
covered n zones and the total cannot be split between them.