How these benchmarks work
Pre-registered, reproducible, fair. Every card's pass criteria were written before any system was run. The THIRI column renders live in your browser from the engine's exact output — the same notes are in the downloadable MIDI and the receipts JSON. The generator column is that system's unedited first result for the identical prompt.
- These are different kinds of products. A text-to-music generator produces finished audio — often beautiful. THIRI produces directed, editable musical structure. The test is not which sounds more produced; it is which one did what the musician asked.
- Audio stems are not parts. Some generators export stems. A stem is a frozen performance — you can turn it down, but you cannot change its ii chord to a tritone sub. MIDI is editable intent.
- Determinism is a feature. Every THIRI result on this page reproduces byte-for-byte on re-run. Generator output changes on every run — by design.
- The produced render is the same MIDI. Where a produced version appears, it is this page's downloadable MIDI dragged into a DAW with instrument sounds chosen — minutes of sound selection, zero notes changed.