sky-truth air-chemistry comparison plan, version 2, written 2026-10-24, before any chemistry was read

History: Version 1 was stamped on 2026-10-24 at about 00:02 Central. After an outside review the same night, and before any chemistry had been downloaded (the database held 0 EPA rows when the job was stopped), it was replaced by this version 2, which adds the fake-date check, ground-wind groups, week-block shuffling, longer delays, detection limits, and waiting for Bitcoin to confirm the plan first.

The program reads no chemistry until Bitcoin has confirmed the stamp of this text.

CHEMISTRY. EPA Chemical Speciation Network, Dallas Hinton (AQS 48-113-0069), PM2.5 elements by X-ray fluorescence,
24-hour samples (midnight to midnight local), sample by sample from EPA's AQS API with EPA's own detection
limit for each sample. For each element, the sampler position (POC) with the most samples is used. A value
below its detection limit is set to half that limit. An element is TESTABLE only if at least
50% of its samples on labelled days are at or above the detection limit; otherwise it is
reported as "not testable" with its detected share, and given no result.

DAY LABELS (weather only, never chemistry). NOAA GFS analyses every 3 hours at 200, 250, 300, 350 hPa
over the city. A time has lasting-trail air when the Schmidt-Appleman test (engine efficiency 0.35) says a
trail forms and humidity is at least 100% over ice at any of those levels. Over the sample day and the day
before (16 times): lasting-trail day = 4 or more times; no-trail day = none;
anything between is left out. A day with any time missing from the archive gets no label. From 2021-04-01.
Every calendar day is labelled, not only sample days.

GROUPS. Days are compared only with days in the same calendar month (any year), the same rain group
(>= 1.0 mm on the sample day or not) and the same ground-wind group (dominant 10 m wind from the south
half, 90 to 270 degrees, or not). Rain and wind: ERA5 via Open-Meteo.

PRIMARY MEASURES: aluminum (88104), barium (88107), strontium (88168), each divided by silicon (88165) to take out wind-blown dust.
STATISTIC, one for each: difference in mean log(element / silicon), lasting-trail days minus no-trail days,
within each group, weighted by n1*n2/(n1+n2) and averaged over groups.

P-VALUE. Weather patterns last several days, so whole weeks (Monday to Sunday) are shuffled, not single days:
20000 times (seed 20261024), the label series of each week is moved to another week from the same
calendar month, and the statistic is recomputed. One-sided p = share of shuffles at least as large as the
real one. Holm correction across the three. A finding needs adjusted p < 0.05 and at least 30
lasting-trail and 30 no-trail sample days.

FAKE-DATE CHECK. The identical test is run with labels taken from 7 and 14 days LATER,
which cannot have caused anything. If either fake-date test also gives p < 0.05 for an element in the same
direction, that element's result is reported as "not separable from weather and season", not as a finding.

SECONDARY, clearly marked: labels from 3 to 10 days EARLIER (slow fall-out; Holm across all of these);
raw concentrations without the dust correction; and every XRF element (Holm across all of them).

WHAT A RESULT CAN AND CANNOT SHOW. Particles released at cruise height mostly take days to weeks to reach the
ground and land spread over hundreds of miles, and about the same number of planes fly over Dallas every day:
a lasting-trail day means the upper air let trails last, not that more planes flew. A finding (that passes the
fake-date check) would mean more of an element at ground level after lasting-trail weather than after dry
weather, beyond dust, rain, wind and season; it would not by itself show where the element came from. No
finding would mean this monitor shows no such difference; it would not rule out releases whose fall-out lands
elsewhere or is too small for the monitor to measure. Nothing here is changed after the chemistry is seen.