Study 4: are the civilisers one story?

The argument goes back at least to Daniel Brinton's American Hero-Myths (1882), which set Quetzalcoatl, Viracocha, Itzamna and Michabo side by side, and it is central to Graham Hancock's Fingerprints of the Gods: bearded, pale strangers arrive by sea after a catastrophe, teach agriculture, law, building, writing and the calendar, are gentle, forbid sacrifice, are driven out or murdered, and leave promising to return. Hancock reads the agreement as memory of the same people, survivors of a drowned civilisation, remembered in Peru as Viracocha, in Mexico as Quetzalcoatl and Kukulkan, in Colombia as Bochica, in Sumer as Oannes and in Egypt as Osiris. The scholarly reading is that each figure is at home in its own tradition, that the "white and bearded" details in the American cases appear in Spanish colonial sources and served Spanish purposes, and that culture heroes who bring fire, crops and law are among the most widespread figures in world mythology because every society has a story about where its arts came from.

That last point is the test. If civiliser myths are alike everywhere, the Hancock pairs should be no more alike than any two culture-hero stories from unrelated places. If the Hancock pairs share something specific, a judge who cannot see the names should pick each out from a lineup of nine strangers, the way it can pick Genesis out of a lineup when shown Gilgamesh.

Each myth is the narrative of one English Wikipedia article, chosen because it is uniform, cited and reproducible (the revision id is recorded for every one). Claude Sonnet 5 retold the myth in 200 to 300 words using only the article, then rewrote it with every personal, divine, place, people and language-specific name replaced by a role: the hero, the sky god, a great lake, a foreign people, a grain crop. Actions, sequence, objects, animals by kind and numbers were kept. Below, Viracocha before and after.

Does the blinding hold? No. A judge that recognises a story is not blind. Claude Opus 5 was shown every de-identified text and asked to name the region and the figure. It named the region correctly for of stories, said it could not tell for , and was right with high confidence for of the it was confident about. That leak is reported for every target below, and it changes what the lineups measure: not a blind reading of two texts but the model's considered judgment of two stories it knows, with the names removed. The lineups are kept because that judgment is still worth having, with the controls to calibrate it. Two wording-based comparisons are reported in the next section; neither solves the source-selection problem.

The same texts were compared with a local sentence-embedding model (all-MiniLM-L6-v2) and TF-IDF word overlap. Each partner is ranked against the other stories by cosine similarity. These comparisons are sensitive to wording, length and the preparation of the summaries. A pretrained embedding model is not knowledge-free: an inability to name a myth does not establish an absence of learned associations. TF-IDF avoids that particular issue but still inherits the source and rewriting choices.

One target story and ten candidates, with their names removed. One candidate is the planted partner; the other nine are random culture-hero stories from regions other than the target's and the partner's. The judge ranks all ten by story content: roles, actions, sequence, what is brought or taught, the specific motifs, weighing an odd shared episode far above "a god, a journey, a punishment". If the planted story is nothing special it ranks first one time in ten and averages 5.5; if it stands out it ranks first most of the time. Each pair is run both ways with ten different decoy draws, by two judges, Claude Opus 5 and Claude Sonnet 5. The code also excludes figures used in the named pair sets from the decoy pool. Decoys are not tightly matched on teaching role or source coverage. The repeated lineups reuse stories; they are not independent historical cases.

Four sets. The civilisers: the pairs the claim rests on, Viracocha with Quetzalcoatl, Oannes and Osiris, Quetzalcoatl with Oannes and Osiris, Oannes with Osiris, and Bochica, Votan, Thoth and Nommo with their supposed counterparts. Documented transmissions, the yardstick: the Gilgamesh flood and the Genesis flood, Atra-Hasis and Gilgamesh, Utnapishtim and Noah, the Ramayana with the Khmer Reamker and the Thai Ramakien, the life of the Buddha and the Christian legend of Barlaam and Josaphat that was made from it, the Panchatantra and its Arabic translation Kalila wa-Dimna, Manu and Matsya, Deucalion and Noah. Related by contact: Quetzalcoatl and Kukulkan, Viracocha and Manco Cápac, Osiris and Isis, Enki and Oannes, Fuxi and Nüwa, the Māori and Hawaiian Māui, Idris and Enoch. Random: a random target and a random partner from another region, the base rate.

For each civiliser pair, the random decoys that were ranked above the planted partner most often. These are candidates for further inspection, not proof that a pair lacks specific correspondences. Rankings depend on the available alternatives and the judge.

Every story, its region, its source article and revision, and its de-identified text. Click a title to expand.

Wikipedia is a secondary source. Its articles compress and select, and for the American civilisers they already carry the colonial-era framing (bearded, white, from the sea) that is itself in question. A full study retells from the primary sources: Betanzos, Cieza de León and Sarmiento for Viracocha; Sahagún and the Anales de Cuauhtitlan for Quetzalcoatl; Berossus for Oannes; Plutarch and the Pyramid Texts for Osiris.

The blinding leaks. The leak test shows that the model recognises many of these accounts. Removing names removes the easy route to recognition, not the hard one, and the leak test shows how often the region could still be named. Where a judge recognises both stories in a pair, its ranking may be its opinion about the pair rather than its reading of two texts. The leak column in the pair table marks those cases.

Two judges from one company, and uneven completion. The models may share biases, and some Sonnet lineups failed. The published counts describe completed trials. More model families could help assess robustness, but cannot make recognisable summaries blind or repair a mismatched comparison pool.

Story similarity is not transmission. Falling below the positive controls does not imply being no more similar than strangers: the claimed pairs were well above the random pairs in this setup. The controls also mix close retellings with broader proposed relationships, so their score is not a universal threshold for common ancestry.

Repeated trials are not independent evidence. The original binomial p-values are retained for transparency, but their trial-independence assumption is not an adequate model of uncertainty across traditions, sources and judges. They should not be read as probabilities of a historical explanation.

What would improve the inference. Trace the shared details into identified source passages, preserve variants and compare against other teaching or founding figures with comparable coverage. The version 2 source audit begins that work. It does not turn this pilot into a test of worldwide rarity.