Study 23b · The chart-draft pipeline
The chartless-song grid collapsed the draft, and the same-day correction relocates the cause: the bench's ruler, not the band's tempo
The chart-draft pipeline, chartless
Study Overview
The Arms
Both arms build the grid a chartless song would actually have: the period from the song's declared bpm and meter, the bar count from the recording's length, the transcription binned by that grid, the shipped reader's labels over that grid as hints, and Study 23's prompt, model and call shape verbatim. B2 anchors the grid at the aligned truth's first bar start, simulating a user who taps the downbeat perfectly once. B1 takes Study 22's blind onset-comb phase, the zero-UX option. Gates sat on the JSON chords channel, index-paired against the chart, with the grid-given baselines recomputed in-run from Study 23's committed results.
Gate One: the Perfect Tap Fails Everywhere
The tap arm was the worst arm measured. Roots against the grid-given baseline: letter-lands 39 of 76 against 72, eat-all-the-clocks 21 of 84 against 68, peaceful-dreams 64 of 153 against 151, mop-heart 53 of 140 against 131. The mechanical anatomy leaves no room for argument: B2's grid hits the truth lattice at exactly one bar per song, its own anchor, and its mean per-bar error runs 3.3 to 8.4 seconds, multiple bars adrift by mid-song. The phase was perfect by construction, so the entire error is the period. The declared bpm is not the recording's tempo, and a uniform grid built from it walks off the music within a dozen bars no matter where it starts.
The Inversion: the Blind Comb Beats the Tap
Gate two passed, but inverted: the blind comb landed above the perfect tap on all four songs (52, 51, 138 and 133 roots), with mop-heart at 133 of 140, above its own grid-given baseline. Two mechanisms, both visible in the committed records, neither a vindication of blind grids. The comb fits the whole recording's onsets, so its wrong-by-whole-bars phase acts as a crude mid-song recentering of the wrong-period lattice, roughly halving the mean error. And these songs' progressions repeat, so index-paired roots forgive a grid shifted by whole loops. B1's grids are still not truth, seven lattice hits at best; the scores survive because the music is periodic, not because the bars are right. The dialect casualty confirms the stakes: peaceful-dreams' 62 chart-dialect bars, eleven exactly right grid-given, drop to four under the tap and six under the comb.
The Shootout and the Decision
The $0 side-measurement answered the tap-free downbeat question: graded by circular anchor error against the truth's first bar start, the onset comb hits three of four songs within 15 percent of a bar, the chord-change comb two, the bass comb one, and the music-start anchor two by circular error. A correction, dated the same day and found by Study 24's variant data: this entry first said the music-start anchor caught letter-lands' 7.15 second intro at 121 ms, and it did not. Its anchor there is t equals zero, the recording opens with a loud note four bars before the charted music, and the 121 ms figure is mod-period coincidence, exactly what a comb would report; across the corpus the first-loud-note anchor finds the audio start, not the charted start, so nothing measured yet detects an uncharted intro. So anchoring within the bar is nearly solved and it is the wrong question: B2 just proved a perfect anchor is worth nothing under a wrong period. The registered decision table fires its third row: uniform grids are insufficient even anchored, and the feature needs alignment-grade grids or the true tempo measured from the recording. The cheapest candidate reuses the shipped reader with zero new UX, extending the chord-change comb from a phase search to a joint phase-and-period search, and is registered as the program's next study. The prediction scorecard is honest about the surprise: the gate-one failure was called but not its universality, and the comb-beats-tap inversion was predicted in the opposite direction.
Correction: the Tempo Was Never the Problem
Published the same day, prompted by Kevin, verified off the raw stems mix with no alignment involved: a wide fine comb sweep over the onset envelope finds razor-sharp peaks at 131.97 to 131.99, 120.00, 144.96 to 144.98 and 223.82 to 223.97 bpm against the declared 132, 120, 145 and 224. The recordings are click-exact. Three findings replace the withdrawn verdict. First, the alignment truth's bar geometry is unusable far beyond its known 400 ms wander: its bar durations on letter-lands span 0.91 to 2.43 seconds against a true 1.818, though its per-bar bins stay musically aligned, which is why the program's chord grades stand. Second, the 7.15 and 6.5 second 'intros' on letter-lands and peaceful-dreams never existed: both recordings open on chart bar one at t near zero, the ruler's start was four bars late, and the 'perfect tap' arm inherited exactly that error, which also supersedes this entry's earlier music-start correction (the audio start IS the charted start on these songs). Third, eat-all-the-clocks' chart needs 172 seconds at its declared tempo and the recording is 147.5: about 13 chart bars have no audio, and no grid of any kind can chart unplayed bars. Re-scored against the true periods, the follow-up study's onset joint comb had found the tempo on all four songs to 0.02 percent or better, and with the true grid the shipped reader's hints alone reach 70 of 76, 134 of 153 and 111 of 140 roots on the arrangement-matched songs. The corrected guidance to the app: use the declared bpm, it is exact; anchor bar one near t equals zero; the open problem is arrangement mismatch, not tempo.