Downbeat disambiguation fails both gates: chroma novelty is not a downbeat detector, and bar one turns out to be a bit the user should supply

The chart-draft pipeline · the disambiguation study

Study Overview

Two Scorers, Registered in Advance

The candidates were the Study 28 anchor's whole-beat shifts, every distinct lattice position in one bar, each kept inside the physicality window. The scorer was the mean detrended energy at all bar lines of the candidate grid. D1, the registered primary, used chroma novelty, one minus the cosine similarity of adjacent chromagram frames, on the theory that harmony moves on downbeats and the backbeat carries drums, not chord changes. D2, the reported secondary, used the onset envelope of the mix low-passed at 150 Hz, the kick band, with the registered expectation that kick-on-one-and-three patterns would leave a half-bar ambiguity in 4/4. The registered prediction was that D1 would pass on all four songs. It passed on none.

The Anatomy of the Zero

The committed candidate scores tell the story plainly. letter-lands' backbeat outscores its true downbeat. mop-heart's two leading candidates tie to four decimal places. peaceful-dreams' winner leads by four percent, inside the noise. And eat-all-the-clocks, the one song Study 28 had anchored correctly, was moved off its correct base to a wrong shift. The mechanism is an old acquaintance: the chromagram's long analysis windows smear harmonic content across neighboring beats, the same smear that made the DTW's interior bar geometry wander in Study 26, so a frame-to-frame novelty signal has no beat-level resolution left to offer. The kick band did better, right on letter-lands and mop-heart, but kept peaceful-dreams' backbeat and also derailed eat. The re-drafts closed the loop: letter repeated its 81.3 percent and peaceful fell to 61.4, worse than Study 28, because its anchor was now wrong in a new way.

One Bit, and Whose It Is

The useful product of three studies is a precise division of labor. The bench has now measured three cheap chartless downbeat scorers, raw comb phase in Study 28 and chroma novelty and kick-band onsets here, and on every song at least one of them preferred a wrong shift over a correct base, so no measured signal earns the right to overrule the anchor it was asked to check. Meanwhile the rest of the chartless grid is machine-exact: the period to 0.02 percent, the bar count within one of the chart, the lattice itself right. The entire remaining uncertainty is one bit per song, which beat is bar one. The named remedy, delivered to the app, is to collect that bit from the user as a single bar-one tap, not a tempo tap, since the grid is already exact, while the alignment feature ships v5 unchanged because its DTW tether supplies the same bit mechanically. Two DSP candidates stay on the table unregistered, bass-note changes at candidate bar lines and hint-change concentration, if a zero-tap answer is ever worth another study.

Correction 4 (2026-11-08)

Re-read against the corrected eat anchor and by beat class (Study 31c showed the class is what transfers, with the whole-bar representative placed separately): chroma novelty was 2 of 4 and the kick band 3 of 4, not 0 and 2. Novelty's eat and peaceful choices were the right class; peaceful failed only the absolute tolerance plus the whole-bar placement defect Study 31c fixed, and eat's base was not correct to begin with, so nothing was moved off a correct base there. Letter's anatomy stands: its backbeat genuinely outscores its downbeat on novelty. The one-tap remedy this entry closed with is superseded: Study 31b's witness vote finds all four downbeats, and harmonic rhythm alone is 4 of 4.