Each version of a piece below was written live, one bar at a time, by the same music generation model writer, from the same plan, primer and seed, with the same checks. Only the choice of each bar differs: d1-ultra, Jev 1.14, a random pick from the same menu, or no chooser at all. A fifth version is the music generation model on its own, with no checks either.
The writer asks the music generation model for many drafts of the next bar at high temperature, throws out what fails its checks, and offers what is left as a short menu. Code does the first three steps for every version. The four versions differ only in the last step. The music generation model alone, the fifth, skips steps 3 to 5: each bar is one draft that parses, drawn at random.
Pick a piece. The five versions scroll together, and the tinted band marks the same bar in each. Press play to hear one version. Switch versions with Hear at any bar, or click any bar of any version to hear that version from there. Where a section starts with a small square, the four checked versions were offered the same drafts and menu at its first bar, so only the choice differs there.
Each condition wrote the same 204 pieces: six plans, 34 seeds each (the pieces above are the first four seeds of each plan). The panel compared pieces from the same plan and seed in pairs, both ways round. The Elo scale is anchored at Jev 1.14 = 1000, with 95% intervals clustered by piece. The other five decision models are drawn faintly for context.
Share of all its verdicts each condition won within a plan, against whichever opponents it met there.