Findings · in plain language
What we have found
Every number on this page is regenerated from the data indexes at build time. Nothing here is typed in by hand.
The question
Strangers telling the same stories
People who never met — separated by oceans and thousands of years — told strangely similar stories. A hero descends into the land of the dead and returns. A great flood wipes the world clean. A mother goddess loses and finds her child. Why?
There are three boring explanations: the stories were inherited from common ancestors, traded along contact routes, or the similarity is coincidence. And there is one profound possibility, suggested a century ago by Carl Jung and Joseph Campbell: that these patterns rise from something shared in the human mind itself. They could never test it. We can.
What we built
An atlas with receipts
We collected 338 complete ancient and sacred texts from 106 traditions — all public domain, all with provenance — and had AI read every passage and tag each recurring story-pattern (a motif) with the exact quote as evidence. Nothing is asserted without a passage you can click and read in the book. So far: 140,297 distinct motifs across 321,766 tagged occurrences.
Crucially, the corpus includes traditions that could not have borrowed from each other — Australian Aboriginal Dreamtime tales, Inuit stories from Greenland sledge journeys, San narratives from the Kalahari, Hopi and Zuni tellings, Guiana Amerindian legends. When the same pattern appears there and in Gilgamesh, “they copied it” is off the table.
338 texts
106 traditions
140,297 motifs
321,766 evidence-anchored occurrences
Replication
Do different AIs see the same patterns?
We had three different AI models — built at different times, with no shared memory — each read the same ancient passages independently and tag the patterns they saw. If the patterns were imaginary, the three readers would disagree wildly. They don't.
Across 6,771 shared passages, any two readers agreed on 36–71% of the pattern families they tagged (a strict overlap measure where 100% means identical answers). For open-ended reading tasks, that is the level of agreement trained human researchers reach.
Why it matters: it means the patterns live in the texts themselves — not in one reader's imagination.
3 independent readers
6,771 shared passages
36–71% pairwise agreement
Clusters
One web, not many
Jung claimed symbols travel in stable families. We joined motif families wherever they recur together across unrelated traditions, expecting archipelagos — a hero-cluster here, a creation-cluster there. Instead the 346 conserved edges close into a single connected constellation of 28 families. Across 200 shuffles, chance produced 325 conserved edges on average — never 346 (z = 6.81, p = 0.00498).
The full web is drawn on the title page and walkable from the descent.
The preregistered test
The isolated lineages already knew
The sharpest test we have run — and we wrote the prediction down before looking. Learn the pattern-web from 84 connected Old World traditions (40,933 records, 334 conserved pairings), then open 23 held-out lineages with little or no historical contact with that world — amazonian, andamanese, australian aboriginal, guiana amerindian, hawaiian, hopi, and others.
The isolated lineages reproduced 82.9% of the trained web, where chance manages 80.3% and the luckiest of 200 shuffles ever reached 84.4%. The observed rate beat 192/200 permutations — a lift of ×1.033 over chance. The preregistered verdict: leakage_flag
Why it matters: inheritance and trade cannot explain a pattern-web that re-forms among peoples who never touched it. Either the world's stories share deep structure, or chance got luckier than it managed in 200 tries.
82.9% reproduced by isolated lineages
80.3% expected by chance
192/200 permutations beaten
×1.033 lift over the null
The capstone · assimilation-immune
Two worlds that never met, one architecture
The isolated-lineage test had one fair objection: to compare the isolated peoples with the connected world, we sorted their motifs into our families — so of course they looked familiar. The crown experiment removes that step entirely. We let the isolated world and the connected world each build their own motif-families from scratch, blind, never seeing each other or our categories — then asked whether the two webs of “which patterns travel together” have the same shape.
At 64 families each, 52.4% of the isolated world's bonds reappear in the connected world's, where chance yields only 36.6%. The match beat 500/500 permutations (p = 0.002), and a second, label-free test found the two webs' structures statistically indistinguishable (degree-KS p = 0.56). The preregistered verdict: STRONG.
Why it matters: no mapping into shared categories can explain a match between two webs that were never allowed to see one another. It is convergence, not identity — nearly half of each world remains its own — but it is more agreement than chance can reach. The full walkthrough, the numbers, and the honest limits are on the Crown.
52.4% bonds in both worlds
36.6% expected by chance
500/500 permutations beaten
0.56 topology-KS p (indistinguishable)
Order · a falsification we keep on the shelf
Campbell's grammar mostly wasn't
Campbell claimed stories don't just share ingredients — they share a sequence: departure, ordeal, return. We tested it. 72 strong before-and-after orderings emerged from the texts, and under within-text permutation 61 of them dissolved — chance writes those orderings on its own. As a universal grammar, the hypothesis failed; 11 lines of order held. We publish the falsification beside the findings, because that is the point.
Structure · how stories are built
The hero's journey fits the world's stories worst
A companion study read 234 stories across 71 traditions for their architecture — arc, grammar, and craft — then scored them against every famous theory. Campbell's monomyth, the one taught in every screenwriting class, fit worst: the best-fitting framework for just one story of 234. Booker's plain plots and the three-act fit best; the real universals are smaller — recognition, interdiction, departure, return. And unlike myth's content, story form does not provably converge across unconnected worlds: it is everywhere the same, but too simple to carry a fingerprint. The full reckoning is on The Shape of Telling.
What would prove us wrong
The exit conditions
If the patterns only appeared in cultures that had contact; if independent AI readers stopped agreeing when texts got unfamiliar; if the “universal” patterns failed the chance test — the profound explanation would lose, and we would say so. One of our own favorite hypotheses already has. The full apparatus — agreement matrices, permutation counts, per-lineage tables — is laid out in the Lab.