Precipitation regime and Temperature regime both rank the 254 corpus cities by a composite distance: each city's monthly rainfall or temperature pattern — magnitude, seasonal shape, how sharply it swings across the year — is reduced to a handful of numbers, standardized, and compared to the query city's own numbers. The smaller the distance, the closer the match; results are a ranked list, not a fixed set.

Terrain regime works differently. Each of three measures — mean elevation, the range of relief nearby, and where the point sits within that relief (valley floor vs. ridge) — is computed live from a small elevation grid around each city rather than read from a stored value, and compared to the query city's own measures via its own tolerance band rather than one combined score. A city only qualifies if all three measures fall within the query's bands; qualifying cities are then ranked within that eligible set.

All three predate Sandbox's own similarity panel and use this older composite-distance / tolerance-band approach rather than the all-conditions-must-hold conjunction method Sandbox uses — see the project documentation's Similarity page for the fuller comparison.

Each of the 254 corpus cities has a Wikipedia article, and its sections are sorted into four themes — History, Environment, Culture, Modern — by their headings. All the sections belonging to one theme are condensed into a short summary of that theme alone, so each city ends up with four independent short profiles rather than one undivided article.

Each summary — plus a fifth, Composite, built by combining all four — is converted into a numerical embedding, a machine-readable representation of its meaning. Two cities are "similar" under a given band when their embeddings point in a similar direction, regardless of whether they share any actual words — Environment can surface a shared physical setting even when the two Wikipedia articles are worded nothing alike.

This measures how similarly two places are described, not any environmental or cultural fact about them directly — it inherits whatever Wikipedia happens to emphasize or omit for each place, and is a different kind of similarity than the environmental lenses above, which compare measured data rather than text.