Agentic Stemmatics
the philology of machine-generated text

Counterpoint and the AI review panel

Luuk Joseph Soons, 14 October 2026. Drafted with Claude (Anthropic). Reviewed before publication by Kimi (Moonshot AI) and Codex (OpenAI).

Schelling called architecture frozen music. Goethe, crediting the image to “a noble philosopher”, proposed a better one: music that has fallen silent.

The first of the rules of motion in Johann Joseph Fux’s Gradus ad Parnassum (1725) forbids a step that sounds harmless: moving from one perfect consonance to another with both voices travelling in the same direction. Above all, two voices may not travel together in fifths or octaves. One common explanation is perceptual: voices that move together in perfect intervals tend to fuse into one, and the texture loses a line while appearing to keep it. David Huron derived the rule from such principles (2001), as advice for composers who want each voice to be heard as independent. Fux’s title means “steps to Parnassus”, and the book is a staircase: it introduces techniques in stages, checks each by rule before the student moves on, and finally combines them.

A review panel can fail in the same way. Reviewers drawn from one model lineage tend to agree with one another, and the agreement feels like confirmation. It may be a parallel fifth: the voices moved together because they share an ancestor, not because each found the same thing.

Why does this matter? A second opinion is worth asking for only if it could have come out differently. Suppose each of three reviewers misses a given flaw one time in ten. If they miss flaws independently, all three miss it one time in a thousand. If they share the blind spot, all three miss it one time in ten, and their agreement adds nothing: three signatures under one mistake. Three people giving directions from the same out-of-date map will agree with each other, and their agreement tells you about the map, not the road. Independence is what lets agreement count as more than one opinion.

Stemmatics makes the point precise, and adds a condition. Paul Maas called the errors that reveal descent Leitfehler, indicative errors. Agreement in a correct reading proves little, since any careful copyist might reach it alone. Agreement in an error points to a common source, but only if the error is unlikely to have arisen independently. An error that copyists often make on their own is weak evidence of descent, although several shared features together may still be significant. For model panels the condition matters twice. Many shared false positives are mistakes any model might make: they say little about lineage, yet they still make agreement misleading, because the voices fail together. What marks lineage is a shared idiosyncrasy whose independent origin is improbable. So far that has been shown for markers placed on purpose; whether models carry such markers of their own is still an open question. Either way, a panel is worth what its members can fail at differently, and that has to be measured, not assumed from the names of the models.

Counterpoint does not stop at prohibition. It prefers contrary and oblique motion: when one voice rises, another falls or holds still. A panel can be composed the same way, with different questions, different entry points into the material, and different things each reviewer is told to doubt. That may expose different weaknesses. Whether it makes agreement more trustworthy has to be measured, and a finding must still be checked against the evidence.

Bach’s Art of Fugue shows the other half of the discipline. It was printed in 1751, the year after his death, in an order that differs from the autograph. Its last fugue breaks off at bar 239, in a section whose subject begins with the notes B-A-C-H, German note names for B flat, A, C and B natural. A note on the manuscript, usually attributed to Carl Philipp Emanuel Bach, says that the composer died over this fugue, where the name BACH had been placed in the countersubject. That note is the source of the famous story that Bach died while writing it, and it is a witness like any other. Christoph Wolff argued in 1975 that Bach had worked out the end of the fugue in a draft now lost; others have questioned that reconstruction. Editors do not conclude that the work is unreliable. They compare the witnesses, weigh their authority and say what remains unknown. A review report owes its reader the same: which voices found the finding, how independent they were, and where the evidence stops.

The Latin contrapunctus comes from punctus contra punctum: note against note. That is what a review panel is for.

Counterpoint does not avoid consonance; first species admits nothing else. But it accepts only consonance reached by independent voices, and a parallel fifth sounds consonant and is still forbidden. The same holds for a panel. Agreement counts only when it comes from voices that could have disagreed, and only when the evidence, not the vote, still stands behind it. When a panel sounds perfectly consonant, check the voice-leading before trusting the chord.

Sources