# Does it rain more in the middle of the week?

Gauge test of Bell et al. (2008): south-east US, JJA, Tue–Thu vs Sat–Mon rain.
`rule.py` is the pre-registration, sealed as `rule.weekly-rain` at 1f916.ai before
any rain data was fetched. `analyse.py` runs it; `results.json` is its output.

Data: GHCN-Daily by_station files (NOAA NCEI, no key), 64 stations listed in
rule.py, fetched 2026-10-08 into ~/data/weekly-rain (not committed; re-download from
the URL in rule.py). Verdict: UNRESOLVED, D = +3.3 % [−1.4, +8.3].
Not a forecast and not advice.

## Power, measured afterwards (power.py, power.json) — NOT a test

Rotated-weekday SE calendars with a planted midweek excess, sealed verdict rule:
28 summers (1998–2025): planted 10 % → SURVIVES 88.5 %; planted 5 % → 37 %.
66 summers (1960–2025): planted 10 % → 99 %; planted 5 % → 53 %.
A planted effect exactly at the 5 % bar can pass only about half the time by
construction (the rule needs the point estimate ≥ 5 %), so the 5 % rows say
little. The useful number is the interval: half-width about 4.9 % at 28
summers, shrinking as 1/sqrt(summers). Resolving the observed ~3 % (half-width
about 2 %) would need about 28 × (4.9/2)² ≈ 170 summers, which no gauge
record has. A future sealed test should either target a ≥10 % effect or pool
far more independent data (other regions, other seasons), sized this way first.