A daily scan of urban data science: what appeared, what it found, and why it matters.

How this is made

Every morning, seven days a week, an unattended pipeline reads 7 arXiv categories and 96 journals. A journal article enters because a journal we track published it. An arXiv preprint enters on a classifier's probability, above a floor set from labelled examples. Neither path is asked to fill a quota, which is why some days are short and some are empty.

What we are looking for: work that matters to urban AI and computational urban science, as the Institute for Latent Publics judges it. The examples the floor is set from were labelled here, by hand, paper by paper. The classifier is a way of applying that judgement to more papers than a person can read — it does not form one of its own.

How well that works, measured: taking the ten papers each path ranked highest on a day, precision was 0.60 on the arXiv path, and 0.66 on the journal path. We set ourselves 0.70 and both miss it. Precision holds at the target through about rank 8 and falls after — so the top of a day is better than these figures, and the tail is worse. All of it rests on a small sample judged by hand, so read these as an order of magnitude rather than as figures.

arXiv indexes about three days late, so the newest rows keep filling in after they first appear — in a later issue, never in one already published. Bibliography, authors and links come from the sources, never from a model; a model writes the summaries and the one-line headline, from the abstract and nothing else. Every tag carries a canonical identifier — OpenAlex, ORCID, ROR — so a method named twice is one thing, not two.

What the rows on an issue count. tag shift — how many of the day's papers carry a tag, against that tag's own daily average over the previous 30 days; listed only above 3 papers and 3× the average, and nothing at all until the archive holds seven days to average over. Both halves count the papers we published, so it compares today with our own archive — a shift can mean our collecting changed rather than the field did. It is recomputed from the archive every time the site is built rather than stored in the issue, so an old issue can gain this row, or see its average move, when a day is filled in later inside its 30-day window. The papers an issue published never change; what moves is what we can now compare them against. shares references — pairs of papers citing at least three of the same works, today with today and today with a paper from the previous 90 days; the shared works are named. Sharing references is not sharing a question, and a paper with a long bibliography pairs more easily than a short one. on the same shoulders — a foundational work cited by four or more of the papers a day named. This is the one row that is not counted over the papers we published: it counts every paper the day named, including the ones under Also published today, whose abstracts we could not read. A citation list is complete without an abstract and a tag is not, which is why this row can use the wider set and the others cannot. Four is a measured threshold, not a chosen one — at two the row lists something every day, at six it is empty — and it is the count of papers, not a claim about them: standing on the same older work is not being about the same thing, and it is not the row above counted again. first appearance — an institution on one of today's papers that has not been on a paper we published in the window before this issue. The window is 180 days or the whole archive behind that issue, whichever is shorter, and the row prints the number it actually used; below 60 days of archive there is no row, because at two weeks a fifth of everything looks new and that is a fact about us rather than about the field.

What we stopped showing, and what came back. Three rows — foundational works cited twice in a day, repeated authors, repeated institutions — were empty on every issue for sixteen days running, and measurement said why: an issue publishes about six papers, and only two or three of them arrive with a reference list or an affiliation, because roughly half are preprints and preprints reach us with neither. Each of those rows needed two of the day's papers to share something; the most any of them ever found was one. Widening the window to thirty days does fill them, and changes nothing from one morning to the next — nought to two entries differ between consecutive days — so it would buy the appearance of a finding at the price of saying the same thing every day. Repeated authors is still in exactly that state and still off the page. The other two came back on different terms rather than being restored: the foundational works row was empty because six papers are too few to share one, so it now asks four of every paper the day named instead of two of the ones we published, and the institutions row was replaced by the opposite question — not who repeats, but who has not appeared here before. Everything taken off is still computed and still in the JSON.

All of it is also available as JSON, with the licence and the field meanings written out.