Building a tree from shared error, the assumptions it needs, and where contamination breaks it.
29 March 2027
Building a tree from shared error, the assumptions it needs, and where contamination breaks it.
Last week you learned to name what kind of mistake each disagreement is. This hour turns that classification into a family tree, and shows exactly where that method fails.
The slides do the building#
The deck works one example end to end: five witnesses of a text you have not read, an evidence table, a candidate tree built edge by edge, a competing tree, and the true production log laid alongside the inference at the end. Read it before the workshop; this page adds the two things a slide has no room for.
The assumption the method needs#
Building a tree from shared error rests on one assumption, stated plainly so you can watch for when it fails: a shared error is evidence of shared ancestry only if there is no ordinary way for two copyists to make it independently. A haplography — the eye skipping from one occurrence of a repeated word to a later one — has no everyday mechanism for happening twice, in the same place, in two unrelated copyists. A normalisation does. That is the whole difference between an edge you can trust and an edge that is lying to you, and it is the same distinction week 5 introduced as omission, repetition and substitution.
What contamination looks like from the tree’s side#
The deck’s competing-tree slide shows what happens when that assumption is quietly dropped: two witnesses that share a normalisation, with nothing else in common, look exactly as related as two witnesses that share a haplography — until you ask which kind of agreement each one is. A tree that trusts every shared reading equally will draw an edge the rest of the evidence contradicts. This is not a hypothetical failure mode; it is the trap built into Wednesday’s own error table, and the check at the start of the workshop asks you to catch it before you have seen the answer.
Inference is not the log#
The deck’s reveal slide is not there to check whether you got the “right” answer. A tree built from shared error is a claim about what the evidence supports, and the production log is a record of what actually happened — they can disagree, and when they do, the interesting question is whether your reasoning was still defensible given only what you had. That is the correction note the workshop asks you to write, and it is the habit The Stemma grades: your mark depends on whether your inference was justified by your evidence, not on whether it matches a hidden key.
Same method, different family#
The audio family from weeks 2 and 3 fits this same method, with one warning: copying mechanisms differ across collections, and the deck’s closing slide applies the method to a dub lineage without assuming a tape deck and a copyist fail in the same way. What carries over is the logic — trust an agreement in proportion to how implausible it is by chance — not the specific list of error types.
Before Wednesday#
Watch the deck. Bring your week 5 witnesses and error table, classified by
type. Do not open a pack’s reveal/ folder before the workshop’s Evaluate
stage.
Workshop: Reading the damage backwards
Teaching team
- Dr Wren Halloway — convenor