Rachmaniclaude

In early 2026, the way to get a coding agent to finish a job was to give it something to stop on. Write code until the tests pass. It wasn't perfect, but it kept the model from declaring victory early, and the exercise of getting there made the code better.

So take a completely different task, one people say a language model can't do at all. Claude had never written music. Ask it to compose, and give it a test for the music. That raises the real question: what would that test even be?

Each of these was composed by Claude Opus 4.6 as written notation: a score, with no audio model anywhere. I scored every piece out of 100, blind, without knowing which version I was hearing. My notes from the time are beside each one, typos and all. Where I wrote down a time, it's a link.

There is a public dataset of 254,077 scores from MuseScore, and some of them carry ratings from the people who played them. We measured 42 things about each score, things like how far the melody ranges, how often the harmony moves, and whether a motif ever comes back, and trained a model on the 5,207 pieces with at least ten ratings to tell the better-rated half from the rest.

On pieces it had never seen it was right 61.5% of the time, where guessing gets 51%. That is a real signal and a weak one: star ratings on sheet music are noisy, and almost everything is rated between four and five.

A predicted rating turned out to be a blunt thing to hand a composer. What worked better was a profile: for each of the 42 measurements, where this piece sits among well-rated music, most important first. The composer reads it once, revises once, and writes down what it took and what it refused.

The first blind test was three songs, each heard three ways: the first draft, a revision guided by the model's predicted rating, and a revision guided by the profile. Nine recordings, shuffled, scored without the key.

There were eleven experiments that month. What held up:

• Two instruments, not twenty. The four cello-and-piano pieces scored 85 to 90. Given a free hand, the composer wrote for six to nine parts and scored 60 to 73. My note at the time: “everything happening at once in the same pitch.”
• Less instruction, better music. The chess pieces above came from one sentence of brief. Rigid pipelines, contracts and mandatory targets all made things worse.
• Obeying the test and arguing with it came out the same. Composers told to follow the profile averaged 87.5 on the chess pieces; composers free to refuse it averaged 87.0.

If targets from hobbyists' scores helped, targets from the masters should help more. So we built new profiles from 2,871 canonical scores, sonatas, quartets, art songs, and ran one brief four ways: three composers steered by the masters, one by the old profile. We wrote down five predictions first. Every one favoured the masters.

All five were wrong. The composers hit the numbers: 40 or 41 of the 42 measurements at or above the median on every piece. The old profile won by a distance. Hitting Beethoven's numbers without Beethoven's craft sounds like this.

A test you can pass without getting better is the thing every test is at risk of becoming. After this, the feedback became a suggestion the composer could refuse.

In August we went back and read the whole apparatus before running anything new. The test had been wrong in ways that were easy to check once someone looked:

• It compared duets with symphonies. The reference music ran from 1 to 46 parts and from 4 to 5,661 beats, and nothing was scaled. A 75-beat waltz was being asked for the number of key changes you'd find in a full orchestral score.
• One check could never fire. The test for notes outside an instrument's range had a condition that was always false. It ran in all eleven experiments and caught nothing.
• One check invented errors. At every change of meter it read the previous bar's time signature. On one piece it reported 21 mistakes that weren't there, and the composer was handed them as real.
• It couldn't hear what I heard. Nothing in it measured how the instruments were stacked, which was exactly what was wrong with the orchestral pieces.

So the honest answer to the question at the top: the test never judged the music. It pointed, loosely, and a composer that looked where it pointed often found something worth fixing. What moved the music most was the model. Opus 4.6 was okay at this. The models since have been a lot better.