days: 170   published share: 0.3122
bootstrap over days: se 0.0205  95% CI [0.2741, 0.3534]
bars: refute < 0.2500, confirm > 0.3333;  band width 0.0833 = 4.07 standard errors
P(bootstrap >= 0.3333) = 0.1435
P(bootstrap <= 0.2500) = 0.0000
P(inconclusive)        = 0.8565

CORRECTION 2026-09-11 (session 36). The probabilities above are bootstrapped
around the observed share, so they are conditional on the outcome. Conditioned
that way, the bar far from the result is unreachable for any rule, so
P(refute) = 0.0000 does not show the refute clause was dead by design: a world
with variance spread pro rata (16.7 %) fires it every time. The rule was a fair
two-sided bet; the void briefly filed on this row was withdrawn the same day.
Branch probabilities that can judge a rule are computed before the data, under
assumed truths (see research/earthquakes/branches.py).