+++ title = "Argued Backward from the Resemblance" description = "In 1928 Joseph Bédier audited a field's published family trees and found they had the shape that let editors pick the answer they wanted. Three disciplines that hit genuine underdetermination, and what each does about it." date = 2026-08-18 draft = false [taxonomies] topics = ["comparative-mythology", "textual-criticism", "genealogy"] kinds = ["underdetermination", "descent", "convergence"] [extra] tier = "public" schema_type = "Article" canonical = "" hero = "articles/argued-backward-from-the-resemblance/hero.webp" hero_alt = "Two separate manuscript fascicles revealed as one continuous folded gathering" podcast = "" mechanism = "Where the same evidence is equally well explained by common origin, independent arrival, or later mixing, the explanation gets argued backward from the resemblance itself — and the shape of the model chosen tends to be the shape that licenses the preferred conclusion." symptoms = [ "Two explanations both fit and the choice between them is made on plausibility", "The structure fitted to the data is suspiciously often the one that permits a choice", "Evidence of contact is asserted from the similarity rather than documented separately", "A conclusion is defended by how natural it would be rather than by a route to it", ] apparatus = "Textual criticism, which is the only one of the three to have audited its own published results for the shape bias and named the finding." discriminating_test = "Ask what independent evidence of contact exists — a documented route, a shared error, a dated intermediary — that is not the resemblance itself. If the answer is only that convergence seems likely, the question is open, not settled." false_friend = "A resemblance with a genuinely documented contact route. These exist and are not underdetermined; the Mesopotamia-to-Genesis flood link has shared textual and political history behind it." test = "Audit a corpus of published attribution or lineage claims for structural shape. Bédier's method was to count branches. The finance analogue is to count how often a fitted model has exactly the number of factors that makes the preferred conclusion available." [[extra.faq]] q = "What was Bédier's finding about family trees of manuscripts?" a = "Reviewing stemmas published by editors working on unrelated medieval French texts, he found an overwhelming and statistically suspicious tendency toward exactly two branches. His argument was that a two-branch tree leaves the editor free to choose whichever reading they prefer whenever the branches disagree, so editors were unconsciously building the structure that granted them that freedom rather than the one the evidence supported." [[extra.faq]] q = "Why can't similarity alone establish that two things share an origin?" a = "Because resemblance is equally consistent with shared inheritance, transmission between them, and independent arrival at the same solution. Choosing among the three requires evidence beyond the resemblance. Julian Steward made the general form of this argument in 1929: both diffusionist and independent-invention accounts tend to be argued backward from the similarity rather than from separate evidence of contact or of convergent conditions." [[extra.faq]] q = "What is contamination and why does it break the method?" a = "The stemmatic method assumes each copy has exactly one parent, which is what makes a shared error indicate shared ancestry. A scribe blending readings from more than one exemplar breaks that assumption, and once the blending is pervasive, shared errors stop separating witnesses into a tree at all. The equivalent in any lineage is a node with several unrecorded inputs." [[extra.faq]] q = "Can every individual record be sound while the conclusion is wrong?" a = "Yes. In genealogy an undocumented non-biological parentage event produces a documentary record that is internally consistent and well sourced, because the people creating it had no reason to record anything else. The conflict surfaces only when genetic evidence is correlated against the whole set." +++ In 1928 Joseph Bédier went looking at other people's family trees. Not human ones. Stemmas — the diagrams textual editors draw to show how surviving manuscripts descend from a lost original. He collected published stemmas from editors working on unrelated medieval French texts, people with no connection to each other and no shared agenda, and counted the branches. Far too many had exactly two. There is no reason a manuscript tradition should bifurcate. Some do, some don't, and across a large enough sample the distribution should look like whatever the messy history of copying produces. What Bédier found instead was a strong tendency toward the one shape with a specific property: **when a stemma has two branches and they disagree, nothing in the method tells you which to follow.** The editor chooses. His conclusion was that editors were unconsciously building the structure that returned their freedom to them. Nobody was lying. Every stemma had been constructed from real evidence by a competent scholar. The bias was in which of several defensible trees got drawn, and it ran in the direction that left the drawer in charge. ## Three fields, one wall Bédier's finding is the sharpest instance of something all three of these disciplines run into. **Comparative mythology** hits it first and worst. Two traditions share a motif. That resemblance is consistent with shared inheritance from a common ancestor, with transmission from one to the other, and with the two arriving at it separately — and picking among them takes evidence that is not the resemblance. Julian Steward wrote the general form of this in *American Anthropologist* in 1929, a year after Bédier. Both diffusionist and independent-invention explanations, pursued as general theories, get argued backward from the similarity rather than forward from separate evidence of contact or of the conditions that would produce convergence. Which lets either theory account for the same data, permanently. The flood-myth case shows both states cleanly. The Mesopotamia-to-Genesis link has an independently documented contact route — shared Near Eastern textual and political history — so diffusion there is supported by something other than the resemblance. The wider global distribution of flood narratives has no comparable documented route, and the independent-invention account rests on the plausibility that river-valley and coastal societies would each produce flood stories. That is a plausibility argument. It may well be right. It is not a demonstration, and the distinction between the two halves of that example is the whole skill. **Textual criticism** has the sharpest tool and knows exactly where it fails. Shared error indicates shared ancestry, because nobody invents the same mistake twice. That works only while each witness has one parent. A scribe blending readings from two exemplars — contamination — breaks the premise, and once it is pervasive, conjunctive errors stop sorting manuscripts into a tree at all. W. W. Greg added a quieter problem: the method assumes an error is always distinguishable from a correct reading, which he called wholly unwarranted. A scribe who correctly fixes a predecessor's error produces a reading that looks original *because it reads better* — erasing the evidence an editor would need to detect the correction. The successful repair is invisible by construction. **Genealogy** hits the version where the records are all fine. An undocumented non-biological parentage event leaves a paper trail that is internally consistent and well sourced, because everyone recording it had no reason to record anything else. Every document passes scrutiny individually. The conflict appears only when DNA is correlated against the entire set, and the base rate sits somewhere around 2 to 4% depending on whose review you take. ## What they do about it None of the three solved it. What they did instead is worth more than a solution. They named the state. A stemma is treated as a defeasible hypothesis rather than a result. Comparative mythology's own reference literature states plainly that the method cannot, on its own, separate heritage from diffusion from invention in most individual cases. Genealogy writes conclusions at a form matching their complexity — the write-up's shape is itself a declaration of how strong the claim is. And Bédier's contribution was to audit the field's *published output* for structural shape, which is a different act from checking any individual result. Every stemma he looked at was defensible. The pattern across them was not. ## The same audit, unrun Two firms hold the same position, built on the same model, carrying the same error. Convergence, transmission, or common source. The evidence usually does not decide, and the resolution offered is normally a plausibility argument — these are smart people looking at the same data, of course they arrived at the same place. Which is Steward's move exactly: reasoning backward from the resemblance to the mechanism that would explain it. But Bédier's version is the one nobody has run. Take a body of published financial research and audit it for structural shape rather than for correctness. How often does a fitted model carry exactly the number of factors that makes a conclusion available? How often does an attribution analysis decompose into precisely the buckets that permit the preferred story? How often does a lineage reconstruction terminate in exactly the two candidate sources between which the analyst is then free to choose? Each individual piece will be defensible. That was true of the stemmas too. The finding, if there is one, lives in the distribution — and the distribution is the thing nobody looks at, because looking at it requires suspecting a whole field of a bias that no member of it committed on purpose. Bédier did that to his own discipline, in print, at the height of it. The stemmas did not stop being drawn. They stopped being trusted quite so far.