Textual criticism

How can scholars reconstruct a lost original from surviving copies that all disagree with each other?

The author's manuscript is gone. What survives are copies of copies, made by hand over centuries, and no two of them read alike. One has a word the others lack; another has a whole sentence nobody else does; a third makes better sense in a passage where the rest are gibberish.

The instinct is to despair, or else to count — to take whatever most manuscripts say. Both are wrong, and the second is wrong in an interesting way. Before we get to why, ask the question that decides everything: if the original is unavailable, what exactly is the evidence here, and what is it evidence of?

Begin with the only mechanism we can be sure of. A scribe copying by hand makes mistakes — eye skips from one line to another ending in the same word, an abbreviation is misread, a marginal note wanders into the text. Now the crucial part: he copies what is in front of him. So his mistakes are passed on to everyone who copies him, and are absent from everyone who does not descend from him.

Sit with that for a moment. What does it make an error into? Not noise — a marker. Error is inherited, and inheritance leaves a pattern. If two manuscripts share a mistake, either they inherited it from a common ancestor, or they made the identical mistake independently.

So how do we tell those apart? By asking how likely the coincidence is. Two scribes might both misspell a common word, or both drop a repeated syllable — those errors are cheap, and prove nothing. But two scribes independently omitting the same eleven-word clause, or independently inventing the same nonsense name? That is not plausible. Critics call these significant or indicative errors, and the whole edifice rests on them.

Now the move that most people miss, and it is the elegant one. Suppose two manuscripts share a correct reading. What does that tell us about their relationship? Nothing at all — both might have it from the original, however distantly. Shared truth is uninformative about descent; only shared error is informative. The evidence in this discipline lives entirely in the mistakes.

With that, a family tree can be built from the bottom up. Group the witnesses by the significant errors they share; the groups nest; the nesting is a genealogy. Two rules fall out immediately. If manuscript B contains every error of manuscript A plus errors of its own, and A survives, then B is a copy of A and carries no independent testimony — so it is struck from the calculation entirely, however beautiful it is. And once the tree exists, the branches vote, not the books. Fifty manuscripts descending from one careless ancestor are fifty copies of a single opinion. Witnesses are weighed, not counted.

Then, at each disputed point, read off the tree: whichever reading is attested by the greater number of independent branches was in their common ancestor. Where two branches split evenly, the tree falls silent and judgement takes over — most usefully the principle that the harder reading is likelier to be original, since a scribe faced with something odd tends to smooth it into sense rather than roughen sense into oddity.

Two honest limits. The method assumes each copy had one exemplar. Scribes sometimes had two, correcting one manuscript against another, and this contamination scrambles the tree into a network — it is the reason stemmatics works far better for some traditions than others. And notice what the reconstruction actually recovers: the archetype, meaning the latest common ancestor of the surviving witnesses. That may sit generations below the author. Nothing in the method reaches above it.

The instinct to take the majority reading now looks like what it is: a category error. It treats manuscripts as independent observers when most of them are relatives, and a text's popularity in the fifteenth century says nothing about its fidelity to the second.

It is also worth naming a real controversy rather than presenting the method as settled. Joseph Bedier observed in the 1920s that published stemmata come out bipartite — splitting into exactly two branches at the top — far more often than chance should allow, and argued that this reveals the editor's own hand, since a two-branch tree conveniently leaves the deciding vote to the editor's taste. The objection has never been fully answered. Responses today lean on computational phylogenetics borrowed from biology, and on methods such as the Coherence-Based Genealogical Method used for the Greek New Testament, which build relationships between individual readings rather than assuming a single clean tree of manuscripts.

• How editors of heavily contaminated traditions proceed when no stemma can be drawn at all.
• What phylogenetic software borrowed from evolutionary biology adds, and what it quietly assumes.
• The opposite school, which prints one good manuscript as it stands rather than reconstructing anything.

Every term the collection defines is gathered in the glossary.