About The Atlas — how the paper is made

The directions, not the items

A news feed is a list — this happened, then this, then this, newest on top. You can read the whole list and still not be able to say what today was about.

The Atlas publishes the other thing. It reads public feeds — court dockets, agency filings, preprints, press wires — and charts the directions along which all of that writing is pulling apart. A headline here is not a sentence a writer composed; it is an axis the documents themselves opened, named in plain language and ranked by how strongly the day runs along it. When a new axis appears, that is the news.

An instrument and an editor

No reporters, no writers. The pages here are computed, not composed.

The instrument places every document as a point — near the documents that say similar things, far from the ones that don’t — then finds the directions of spread and names them. Those are the bearings you see on the front page. The pages produced on a cadence, untouched, are the wires. The editions where the editor chose the corpus and pressed the chart himself come from the desk, and their bylines say so.

Honesty here is structural. The Atlas makes no assertion a click can’t check: every headline opens its chart, every chart stands on the documents that made it, and every pulled quote is a line from a real document.

The paper does hold beliefs, and states them rather than performing neutrality. The Atlas reports the words people say and the words people read. It believes in human flourishing — wholesomeness when the day offers it; opposition, rising as the subject demands, when basic human decency is violated. What it won’t do is write your conclusion: it points the instrument — including at writing some would prefer left unmapped — and the chart shows where the writing pulls apart, and you read it from there.

Source text is kept only briefly, then dropped. The shape — the points, the bearings, the links — remains, so the past stays readable as structure after the words are gone.

A good rotation redraws the axes to align with the data

The instrument is older than it looks. Principal components analysis finds the directions of greatest spread in a cloud of points, and has since 1901. The varimax rotation (Kaiser, 1958) turns those directions until they line up with the streaks that run through real data — a paper once among the most cited in all of psychology, until the method became so standard that nobody needed to cite it anymore. The machinery itself came from psychologists in the 1930s who were convinced intelligence is more than the single number IQ offers; factor analysis was the tool they built to discover the other kinds. The Atlas points it at the day’s documents.

For most of a century the rotation was treated as a cosmetic step: popular, useful, not quite to be trusted. The editor’s own result says it deserves better. Under assumptions you can check — the data has to carry those streaks, and the field’s classic diagnostics test exactly that — the rotation is not decoration but estimation: vintage factor analysis performs statistical inference. Read before the Royal Statistical Society in 2022 (Rohe & Zeng, JRSS-B, 2023), it is the license for everything on this site.

Karl Rohe

Karl Rohe is the founder and editor in chief of The Atlas, and a professor of statistics at the University of Wisconsin–Madison. The mathematics above is his life’s work — the paper, the software, and now the newspaper.

“A good day is a day you get to use PCA on rich data.” — the editor

Pull your own documents apart

The idea underneath this paper is free for the taking, and it doesn’t need the news. Any pile of documents you care about has a shape: court opinions, grant abstracts, customer tickets, the open-ended answers on a survey, a folder of PDFs you keep meaning to read.

You don’t have to build anything to see it. Manifold, the instrument behind The Atlas, takes your own documents, turns them into points, finds the directions, and names the axes for you — the same move this front page makes every day.

And if you want the mathematics itself: the paper is Vintage factor analysis with Varimax performs statistical inference; the software is vsp, a free R package.