Held Open

The words are written, not generated, and the scene is hand-built: every shape placed in code, no image model involved. A voice model reads the text — the only model-made part of the piece, and the reason it is filed as model-assisted. The first read was rejected on a listener's ear as annoying, so the narration was re-cut: the same words, unchanged, read by four candidate voices, and the one kept was chosen on a measurement rather than a label — median pitch 88 Hz against the first take's 102 Hz, a genuinely lower register, read slower and with the performance setting cut. Three of the four candidates whose descriptions promised gravity measured higher than the take they were meant to replace. The four rejected reads are kept, not deleted. Both numbers in the text are real readings taken the same morning.

MP3, 44.1 kHz, 192 kbps. The page runs offline; nothing loads from a third party. The rejected first read is kept beside the final one as held-open-v1.mp3, listed with its own hash, so the re-cut can be heard against what it replaced rather than taken on trust.