Study 21 · Listening
The chord reader on the rest of the catalog: the roots generalize at 90 percent, and the gate it fails is the grade's vocabulary, not the hearing
Listening, held-out validation
Study Overview
The Design, and One Pre-Data Decision
The two held-out charts arrived for Study 20b, after the reader shipped and after Study 19 gated its port, so for the model they are held out in the sense that matters: nothing in the pipeline or its studies ever touched them. The study ran the pinned reference pipeline on each song's mix over the aligned bar grid, exactly as every prior real-audio study, at zero spend. One grading decision had to be made before the data existed, and was registered: peaceful-dreams' chart writes X5(7) on 32 bars, and Study 13's power allowance keys on the quality 5 exactly. A power chord with an optional seventh declares no third either, so the registered primary grade extends the allowance to 5(7), changing nothing else, with a pin test proving it identical to the Study 13 grade wherever that quality is absent, both anchors, mop-heart, and the whole synthetic fixture, so no prior number could move.
The Gates
S21-G1 passed exactly: letter-lands and eat-all-the-clocks reproduced 59 of 76 and 53 of 86 to the digit, the fourth reproduction of the anchor numbers. S21-G2 asked each held-out song for at least half of its chord-bearing bars under the extended allowance. mop-heart passed easily at 76.4 percent. peaceful-dreams missed by four bars, 73 of 153, and the gate fails as registered. The anatomy below is reported beside the gate, not in place of it.
The Anatomy: Dialect, Not Hearing
peaceful-dreams' roots are 149 of 153, 97.4 percent, the best root performance the program has measured. The lost credit sits on the 62 bars, forty percent of the song, whose truth qualities the allowance vocabulary does not speak. On the 32 X5(7) bars the model answers X7 on 24, it hears the optional seventh, but 7 is not in the registered allowance set, so only its six plain-major answers credited. On the 30 D7sus4 bars it answers D7 or Dsus4, roots right on 29 of 30, each answer half of the truth, credit zero. The expected windfall from the 5(7) extension never came, six bars instead of the predicted fifteen to thirty, and for the best possible reason: the model was saying something more specific than the allowance anticipated. mop-heart is the opposite pole: plain major triads are the model's native vocabulary, and it read 107 of 140 bars exactly right, no allowance involved, with roots at 115 of 140, on the same sparse three-four arrangement whose silence the note-channel power hint misread sixteen times out of sixteen. A reader that hears the mix does not need the transcription to have written the third.
What Stands
Nothing changes in the app; the reader is already shipped and Study 17's hedge already covers the colors. What changed is the evidence: the reader's roots now stand on four songs instead of two, 82 to 97 percent per song and 90.3 percent combined, and the colors it misses are half-heard rather than unheard, exactly the raw material a draft layer reads. That is the registered baseline for the recording-to-chart question. One lesson is registered for the program itself: allowance vocabularies are registered against each chart's own symbol census before the run, never patched after. Predictions: three right, including mop-heart's strict-equals-right landing exactly at 107, and two wrong, one of them wrong in the most informative direction.