Refused

A tick means that rule would stop that sketch.
Sketch	Pitch	Voices	Controls	Monotony
Phrase 1: Up a fourth, Close	no	no	no	no

Phrase 1: Upside down, Far away	no	no	yes	no

Phrase 1: Wider, Cloud	no	no	yes	no

Phrase 1: Pushed about, Piano	yes	yes	no	no

Phrase 2: Upside down, Cloud	no	no	yes	no

Phrase 2: Wider, Piano	yes	yes	yes	no

Phrase 2: Pushed about, Piano chords	yes	yes	no	no

Phrase 2: Busier, Close	yes	no	yes	no

Phrase 3: Wider, Piano chords	yes	yes	yes	no

Phrase 3: Pushed about, Close	yes	no	no	no

Phrase 3: Busier, Far away	yes	no	yes	no

Phrase 3: Ornamented, Cloud	yes	no	no	no

Phrase 4: Pushed about, Far away	no	no	no	no

Phrase 4: Busier, Cloud	no	no	yes	no

Phrase 4: Ornamented, Piano	yes	yes	no	no

Phrase 4: Reshaped, Piano chords	yes	yes	no	no

Phrase 5: Ornamented, Piano chords	no	yes	no	no

Phrase 5: Reshaped, Close	yes	no	no	no

Phrase 5: Thinned, Far away	no	no	no	yes

Phrase 5: Up a fourth, Cloud	no	no	no	no

The sketches are as they came out, unedited, and mono. The controls come from Grow and the voices from Voice.

What happened

I asked Claude for one crop of sketches from five piano phrases, the first run of what I've been building towards over the last few days.

Claude measured the recordings and found that 12 of the 35 notes it had read weren't really there. The analysis hears a chord as one note, often one nobody played. My brief says nothing is added that I didn't play, so it built a check, stopped 27 sketches, took out controls and voices, and reported 13 sketches and 291 passing tests as done.

I hadn't asked for any of that. Deciding what gets kept is my job, and I said so. I also asked what a monotony check was for, since plenty of music relies on monotony. That rule was already in the project, and Claude had made it stricter. It would have stopped a steady drone.

The twenty above are the second run, with every control and voice put back: all it made, bar one that came out silent.

The misreadings don't bother me. They're part of the exploration. On the Voice page the day before, I had written about the same misreading: “The analysis line of the partly chordal piano phrase, when played back, has an interesting low-pitched artefact. A happy accident.” Claude had built a check to stop my happy accident.

Then I mentioned publishing, and Claude assumed I meant this project and started asking what the site should say about it. They're separate. I publish something every day to be more consistent about my own projects, a bit like a journal.

Exploring, or delivering

AI seems to assume it has to deliver something fully formed. I think that's the wrong aim when you're exploring, and that most people switch off from fully formed AI content and call it slop.

Everything Claude could measure was right here. The tests passed, the audio came out byte-identical on a second run, and the report read like finished work. None of that could tell it the decision wasn't its to make. The polish made that harder to see: 13 tidy sketches looked complete, and hid the 27 it had stopped. I wanted something to react to, and it kept handing me answers.

What I heard

I heard more interesting examples in the crop. I need to think carefully about how to record my thoughts easily, with minimum friction.

What to call it

Up to now I've called this a factory. I think I took the word from software factories, when I was looking for a way past my composition block. I'm not happy with it. It feels more like a continuation of Composition in Slow Motion, my MA Digital Arts final show piece.