Nothing about the weather changes. Only the paperwork does: when the observer reads the thermometer, and what kind of thermometer it is. Watch those two details create, and then erase, hot days in a pristine temperature record.
The US Climate Reference Network (USCRN) is NOAA's gold-standard set of weather stations: triple-redundant sensors, hourly data, sited far from buildings and asphalt, and never subject to the observation-time or instrument changes that ran through the volunteer network during the 20th century. That makes it the perfect laboratory: we can take its clean hourly record for 2005–2025 and replay those historical changes on top of it, to see exactly what they do to a count of days at or above 95°F.
For most of the 20th century, America's daily temperature record was kept by volunteers in NOAA's Cooperative Observer Program, using a pair of special thermometers: one holds the highest temperature reached since it was last reset, the other the lowest. Once a day the observer writes both numbers in a ledger and resets the instruments. Everything hinges on when that reset happens. Reset near the hottest moment of the day, and the thermometer walks into tomorrow's measurement window already remembering this afternoon's heat: if tomorrow turns out cooler, the same hot afternoon gets written down twice. Reset in the morning, near the day's low point, and each afternoon lands in exactly one entry.
This is the time-of-observation bias, and it matters historically because the network moved. In the 1930s roughly 80% of volunteer stations reset in the late afternoon; by the 2000s most had switched to morning readings, largely at the request of the weather services (Vose et al 2003). Every switch was in the same direction, afternoon to morning, so the double-counting habit drained out of the network over decades: a slow, one-way cooling artifact in hot-day counts made purely of paperwork. The chart below shows the mechanism in real hourly data; drag the reset time and watch the ledger change.
USCRN station Manhattan 6 SSW, Kansas · July 3–5, 2023 · hourly maximum air temperature
Over the whole of 2023 this station truly reached 95°F on 26 days. A 5 PM observer's ledger would show 30. A 7 AM observer's: 26.
The second change was hardware. Until the mid-1980s the standard equipment was liquid-in-glass thermometers housed in a white wooden louvered box, the Cotton Region Shelter. Beginning in 1985, NOAA replaced them at thousands of stations with the electronic Maximum–Minimum Temperature System (MMTS): a thermistor in a small plastic radiation shield, read from a digital display indoors. Side-by-side comparisons (Quayle et al 1991; Hubbard and Lin 2006) found the new shields read daily highs around 0.4–0.5°C cooler, mostly because the old wooden boxes heated up in strong sunshine and passed some of that warmth to the thermometer inside. The new sensor wasn't wrong; the old one ran a little hot. But the switch put a step change into thousands of station records at nearly the same time. (We measure this bias directly from 385 documented station transitions in a companion deep dive: The thermometer swap, measured.)
Half a degree sounds small, and for an average it is. A threshold count is merciless to it. A day the old shelter recorded at 95.5°F reads 94.8°F on the new sensor, and simply vanishes from a count of 95°F days, even though the air outside is identical:
The observation-time shift and the instrument swap pushed the same way: both make recent hot days scarcer on paper relative to the past. The panel below replays both transitions on the whole USCRN network at once; set the observation times and the instrument bias, and watch what happens to a count of hot days in data where nothing real changed.
Network-average days ≥95°F per year, 2005–2025 · fixed set of 70 USCRN stations · transitions applied linearly across the period
The instructive part is that the true USCRN record is warming briskly over these two decades: +1.13 days ≥95°F per decade. The simulated observation-time transition alone erases more than half of that (+0.47). The instrument swap alone, at the −0.4°C bias we measured from documented transitions, erases essentially all of it (+0.07). Run both, as the real volunteer network did in the 20th century, and a genuinely warming record flips to apparent cooling (−0.61 days per decade), without a single measurement error anywhere. That is the shape of the bias buried in the raw US temperature record, and it is why homogenization (checking each station against its neighbors for exactly these step changes) is not optional bookkeeping but the difference between measuring the climate and measuring the filing system.
Method. USCRN hourly maximum air temperature (hourly02), 70 CONUS stations with ≥95% complete April–October data in at least 20 of 21 years, 2005–2025. A simulated observer at hour H logs, for each day, the maximum over the 24 hours ending at H (exactly how a max/min thermometer behaves); hour 0 (midnight) reproduces the true calendar-day count. The MMTS transition subtracts a fixed bias from a converted station's readings, equivalent to raising its 95°F threshold. Transitions convert a linearly growing fraction of stations across 2005–2025; historically the observation-time shift took several decades and the MMTS swap under a decade, so this is a stylized compression, not a reconstruction. Code and data: github.com/hausfath/us-hot-days.