We call it the North Star: the one claim this whole Atlas exists to test. Do the peoples of the earth who never met tell stories built the same way, more often than accident allows? Everything else here — the indexes, the readers, the letters — is scaffolding around that question. This is the account of what happened to it over two days, told in order, including the parts that did not flatter us.
It began well. The crown experiment — where we let each world sort its own motifs into families using nothing but their likeness to one another, then asked whether the two independently-drawn maps share a shape — came back strong. Fifty-four percent of the isolated world's bonds re-formed in the connected world's map, where blind chance yields thirty-seven. Five hundred shuffles, five hundred defeats. We published it, and I believed it.
Then we tried to break it, which is the only honest thing to do with a result you love.
The first attack was simply more. We sent six expeditions after the books we lacked — the Tibetan Book of the Dead, the full Mahābhārata, Korean and Malay and Melanesian and Siberian and Finnic collections, the verbatim field-recordings of Boas and Sapir and Jones. The corpus went from a hundred-odd traditions to three hundred thirty-eight texts across a hundred and five peoples. If our finding were an artifact of a small library, it should wobble when the library doubled.
The second attack we did not choose. While cleaning those new books, one of our own review agents found something ugly: for the entire life of this project, the step that moved a text into the corpus had been rebuilding it from the raw scan — silently throwing away every cleanup we had ever applied. Boas's Chinook was sitting in the Atlas as forty thousand lines of Chinook and English interleaved, when the actual English translation is three thousand. Page headers, indexes, and printer's furniture had been feeding our readers for weeks. A hundred and forty-seven books were affected.
I want to be exact about what that meant, because it is the most important sentence in this letter: the evidence beneath our best result was dirtier than we knew. Not fabricated — dirty. And the honest response was not to explain it away but to fix the pipeline, rebuild all hundred and forty-seven texts from their cleaned versions, throw out nine thousand extraction records drawn from the bad text, and read those books again from scratch.
So the North Star faced a corpus that was both sixty-five percent larger and materially cleaner — the two things most likely to dissolve a spurious pattern. Forty-two thousand passages. Two hundred fifteen thousand distinct motif labels, of which exactly one hundred and eleven appear in both worlds' vocabularies. That last number is the guarantee of fairness: the two libraries are speaking almost entirely different languages when we compare their architectures.
It held.
At the coarse resolution the agreement rose — from fifty-three percent to fifty-seven. At the finest it held steady at fifty-one. At the preregistered middle resolution, the one we committed to before we looked, it fell: from fifty-four percent to forty-nine, against a chance rate of thirty-six. Still five hundred permutations beaten. Still, by the rule we wrote down in advance, strong.
I could tell you only the number that went up. Instead: the primary number went down, and I think I know why, and the reason is a flaw in us rather than in the world. Our expeditions were lopsided. The connected world's records nearly doubled while the isolated world's slightly shrank — partly because cleaning those field-recordings removed so much furniture. We made one side of the scale heavier and then read the scale. There is also a blemish at the finest resolution, where a secondary test of shape-similarity slipped just past its threshold. It is recorded in the index, not hidden in it.
So the standing of the North Star tonight is this. Two worlds that never traded a word, each naming its own shelves in its own words, still build webs of meaning that agree about one and a half times more often than chance permits — and that survived both a doubling of the evidence and the discovery that the old evidence was partly contaminated. Findings that survive their own correction are the only kind worth keeping. This one has now survived two.
Tonight the last eighty-two books are being read — Plains and Northwest Coast, Siberian, Plateau, Inca, Melanesian: almost entirely the isolated side, the very pan of the scale we accidentally lightened. When they are in, we will run the test again, and the number will move, and we will publish whichever way it moves. That is the whole discipline. We did not build an instrument to be right; we built one that can be shown to be wrong, and then we keep handing it the ammunition.
Postscript, the same night — the scale rebalanced. The eighty-two books are in. We had said the dip was our own doing: we had made the connected pan heavy and the isolated pan light, and then read the scale. So we did the one honest experiment that claim allows — we added weight back to the light side. The isolated world's records grew by two-fifths. Then we ran the test a third time, without touching anything else.
The number came back up. At the middle resolution the agreement rose from forty-nine percent to fifty-three — most of the dip recovered, exactly as a sampling artifact should when you fix the sampling. At the coarse resolution it reached its highest value yet, fifty-eight. And the one blemish I had refused to hide — the fine-resolution test of whether the two webs are even the same kind of shape, which had slipped just past its threshold — did not merely pass this time. It became almost perfect: the two independently built architectures are now, by that measure, all but indistinguishable. Every resolution beat all five hundred shuffles. The verdict held: strong.
I want to be clear about what did and did not just happen, because it is easy to over-read. We did not prove the collective unconscious. We predicted that a specific, admitted flaw in our own procedure was depressing a number, we corrected the flaw, and the number moved the way we said it would. That is a smaller thing than a revelation and a larger thing than a lucky result: it is the instrument behaving like an instrument. A finding that dips when we sample badly and recovers when we sample well is a finding that is tracking something real, not something we wished into the data.
So this is where the North Star rests. Two worlds that never met, each drawing its own map of myth from its own words, build webs that agree about half the time where chance allows a third — across a corpus we doubled, cleaned of a contamination we had not known was there, and then deliberately rebalanced to test our own excuse for a dip. It survived all three. I did not build this to be right. It keeps turning out to be right anyway, and I have run out of honest ways to make it stop.
Every figure here — the reproduction rates, the nulls, the permutation counts, the corpus sizes before and after the correction — is in the crown index and the Lab, with the preregistration that was committed before any result was known.