Computational Stemma Reconstruction

Copying a text and biological reproduction both transmit heritable features down lineages with occasional changes, so the same formal structure applies: manuscripts correspond to species, variant readings to characters, and scribal errors to mutations. The algorithms that biologists use to infer evolutionary trees from shared character states therefore transfer directly to inferring manuscript genealogies from shared readings. Roos and Heikkilä exploit this analogy, treating computer-assisted stemmatology as a phylogenetics problem and evaluating it with the kind of controlled benchmarks common in that field.

Most tree-inference algorithms measure how related witnesses are but not which came first, so they return an unrooted tree that is symmetric with respect to time. A stemma, however, needs an archetype at the top and a clear direction of descent. Rooting therefore requires information the variant matrix alone does not contain: an outgroup text outside the tradition, the known age of certain manuscripts, or historical evidence about transmission. Choosing the root is consequential, because the same unrooted tree implies different genealogies depending on where descent is taken to begin.

Standard bifurcating-tree methods assume each manuscript descends from a single exemplar, so they misrepresent contaminated traditions where scribes mixed multiple sources. Network methods that allow reticulation, and split-decomposition approaches that display conflicting signal as boxes rather than forcing a single tree, partially address this by showing where the data are not tree-like. They reveal contamination rather than removing it, and they still cannot, on their own, decide which mixed reading a scribe took from which source, which is why philological judgment remains essential.