Study 23 · The chart-draft pipeline
Recording to chart: every gate fails on the chart the app would parse, while the same answers' JSON carries the best chord numbers the program has measured
The chart-draft pipeline, end to end
Study Overview
The Composition
The pipeline was the app's own, end to end: the production transcription path over the aligned bar grid (grid-given, the Study 22 decision), the shipped chord reader's per-bar labels bracketed as hints exactly as production brackets its matcher's, the production draft prompt and call shape, one call per song, about 31 cents against the registered two-dollar fence. Grading ran on two channels of the same answer: the chart markdown read through the production barmark adapter, which is what a user would see and what every gate was registered on, and the bar-exact JSON chords field, production's per-bar channel, reported beside it.
The Gates: Four Failures on the Chart Channel
Peaceful-dreams' chart does not parse at all. Eat-all-the-clocks renders 68 of its 86 bars and mop-heart 84 of 140, so the structure gate passes only on letter-lands, and with bars missing the index-paired root grades collapse. Zero of peaceful-dreams' 62 dialect bars survive on the chart channel. The failure modes are specific and legible: the model writes cell-level repeats, a bar reading D5 x8, which the band's notation does not define and the parser refuses, and it under-writes section repeats. The production prompt's own compact example, D(7)sus4 followed by D5 x2, invited exactly this generalization.
The JSON Channel: the Music Was Right
The same four answers' chords fields are bar-exact on every song and grade as the strongest chord numbers the program has produced. Letter-lands reads 72 of 76 roots, above its hint's 70. Peaceful-dreams reads 151 of 153 roots, above the reader's own 149, with strict credit on 84 bars against the reader's 20, and eleven of the 62 dialect bars exactly right: D7sus4, which appears in the prompt's notation examples, written strictly eleven times where the reader alone scores zero by vocabulary, and G7 on ten G5(7) bars, the seventh heard and the chart's spelling absent from every example, the same half-heard signature Study 21 measured in the reader. Mop-heart reads 131 of 140 roots and 129 strict. The one true regression is eat-all-the-clocks, whose roots drop from the hint's 74 to 68 of 84 on the vocal-heavy song. Had the gates sat on this channel, three of four would have passed; they were registered on the chart channel because the chart is what ships, and they are not re-thresholded.
What Stands
The composition is one deterministic step away from viable. The draft already returns bar-exact chords and section starts as data; a chart built from those fields by code would have passed the validity, structure and color gates on this pilot, leaving only eat-all-the-clocks' root regression to examine. The verdict to the app names that as the recommended design, with constraining the prompt's chart examples to row notation as the weaker second option, and notes that one example line showing the G5(7) spelling would likely convert the ten G7 bars. Any re-test is a new registration. The prediction scorecard is a clean sweep of the right kind: all five predictions wrong on the gated channel, and every predicted mechanism, parse quality, hint-following roots, 7sus4-led dialect recovery, visible in the JSON. The bench predicted the music and failed to predict the serialization.