Where our numbers come from

ERA5 is a grid of roughly 25 km cells: against the 645 stations its annual rain was 1.12 times the measured figure at the median and off by a factor of 1.5 or more for 84 cities, mostly in mountains and on wet tropical coasts; its highs ran 0.7C cooler at the median and were more than 2C off for 81. The 354 ERA5 cities carry that margin; 12 of them are flagged in data/city-climate-source.json because the stations around them show the grid too wet or too dry. A station up to 25 km away can sit in a different microclimate (an inland airport for a coastal town has colder nights), and up to 250 m of height difference is about 1.6C.

They hold no cost components, so the only way to derive a second end was to read a rent range out of their own page prose. Tested against the 330 cities where the true figures are known, that method missed by 23% at the median and by more than 50% on 72 of 488 endpoints, so it was not used: a range nobody measured would look exactly like one we did. The range is also only as good as Numbeo's idea of where the city ends, and it covers rent and a basket, not flights, visas, insurance or health care.

An owner decision to grow volume via aggregator directories took the directory tier from 39 rows to 105, roughly half the set, so the average row is now weaker than when the dataset was 62 rows. Many directory rows rest on a ROSTER-LEVEL claim ("the clinics we list speak English") rather than a statement about the named business; each such row says so in its note. PRIMARY FOR THE CLAIM, NOT FOR FLUENCY. We source that a provider is stated to work in a language; we have not verified that it does, and no row carries the 'visited' tier. A hospital offering interpretation is counted under the languages it interprets, which is not the same as a doctor speaking them: Chiang Mai Ram and Thonglor Pet are interpreter desks, not multilingual clinicians. The 17 directory-tier rows rest on guides and aggregators that may sell placement, making them the weakest entries. The US Consulate Chiang Mai health-facilities and legal-provider PDFs and the US Embassy Georgia mental-health resource PDF are all image-only or protected and could not be read; retrieving those would upgrade several rows and is the obvious next improvement. A Russian-language directory of Russian-speaking doctors in Georgia (needdoc.ru) is an unexploited language-defined source of the same kind. Coverage is a pilot, not a map of any city's real provision, so absence from this list means nothing. REGISTERS (2026-09-24). Japan's register is the facility's own statutory report; only Tokyo's form asks whether a language is accepted without prior consultation, so Tokyo rows require that yes while the other seven cities publish the listed language unless the facility's own note hedges it (translation device, interpreter, call first, a few words). Facilities listing more than six languages and employer clinics are refused. Korean city lists are compilations by a city or district, not statutory reports, and Seoul's pharmacy list is dated 1 January 2023. The Korea Health Industry Development Institute lists language services a hospital offers, which may be an interpreter rather than the doctor.

Speedtest results are user-initiated tests, so they lean toward people who test their connection. Ookla publishes no current mobile median for 39 of the 143 countries and no current fixed median for 2 (Eswatini, Eritrea). Small or new series (Barbados, Eritrea, Turkmenistan: two editions each) pass the same rules but carry more sampling noise.