On 261 songs the detector's bar is right one time in three, a tracker's three in four, and the detector's own 6/8 call is its largest error

Listening

Study Overview

The Corpus

The artist publishes tempo, key, meter and feel on every song page, and sells the whole catalog as one download rather than offering it as a bulk scrape, so the clean path was to buy it. The bench matched 262 bundle files to pages by normalized title (twelve covers and parodies have no original-song page and are excluded), hashed every file and every zip into the committed manifest, and froze the census one commit before the registration: 223 songs in 4/4 (177 straight, 46 swing), 37 in 6/8 (17 straight, 20 swing), one waltz; declared tempos 40 to 280 with a median of 105. The parse reproduces every Study 35 slot value exactly. Truth is the artist's declaration: four quarters in 4/4, two dotted quarters in 6/8, the convention Study 35 checked against the audio. Declared values are the session click, not beat-aligned time truth, so there is no downbeat phase score here.

The Headline Rates

Regular mixes, bar strict within 4 percent, folded at a metrical ratio, and meter class right: production 36, 67 and 69 percent; Beat This final 70, 90 and 85; Beat This small 69, 90 and 87; madmom 66, 89 and 85. On the 177 straight 4/4 songs the trackers reach 73 to 76 strict and 91 to 93 folded against production's 38 and 64, and madmom names the 4/4 class on every one of them. The 95 percent Wilson interval on Beat This's 4/4 strict rate runs 70 to 81 percent on 223 songs, the first tempo number in the series with an interval worth printing. Swing cost the trackers nothing, 76 against 76, where the registration had predicted a ten-point penalty. The registered decision rule for adopting a tracker's bar outright, 80 percent strict and 95 folded, is met by none; the nearest is Beat This at 76 and 94.

The Detector's Own 6/8

Production names 6/8 on 63 of the 223 songs the artist wrote in 4/4, 56 of them straight. Every one arrives through the subdivision test reading three, and 59 of the 63 bars it then builds fold to no metrical ratio of the truth at all. This is the peaceful-dreams confound that Study 11b found and 11c reproduced synthetically, now at catalog scale: the sub-beat test fires on straight running eighths and the ladder constructs a bar the song does not have. The instrumental mixes show it is not the voice (42 of 208 there). When the detector names 6/8 on a song that is in 6/8, 21 of 37, the bar is strict on two. And its tempo confidence does not separate right from wrong on 261 songs: medians 0.762 against 0.685, with twelve right and eleven wrong detections above the 0.84 line Study 34b had read off fifteen recordings. Study 35's advice to the app, do not ship the refusal, stands at scale.

Compound Meter Is Unsolved for Everyone

On the 37 songs in 6/8 the bar is strict for 8 percent of them under production, 32 to 35 under Beat This, 16 under madmom. Beat This reads them as six beats to the bar on eighteen, three on twelve, and lands at the 3/2 fold on eight, a 4/4 bar laid over a 6/8 song. madmom, with bars of three or four only, never satisfies the class. The one song in 3/4 is named by every tracker and read at a fold by all of them. Compound meter is the class the proposal's later tiers (a harmonic-rhythm witness, band-split accents) are for, and this census gives them 37 songs to be graded on.

The Voice Is Not the Problem

The paired arm asked the separation question without running a separator: the same song with and without the vocal. Beat This fixes 10 of its 74 regular-mix failures when the voice is removed and breaks 11 of its 170 passes; the small model fixes 5 and breaks 6; production's outcome changes on 35 of 244 songs, 14 percent, in both directions. The registered threshold for moving source separation up the list was a 50 percent fix rate. At 14 percent, separation goes back down it and the harmonic-rhythm witness comes first.

The Fold Is a Convention, and It Is Merge-Safe

Of Beat This's 39 readings on 4/4 songs that fold but are not strict, 28 are a bar half the declared length and 8 a bar twice it. The half readings sit on songs declared at a median of 72 bpm; the double readings at 212. The trackers live in the 80 to 160 lane, and the artist's slow and fast declarations are session choices outside it. A grid at half the declared bar is a finer grid, which a chart can merge into the coarser one without losing a chord; the coarser direction, where a bar would have to be split, is three and a half times rarer. The proposal's collapse-safe octave policy, prefer the shorter bar and merge identical neighbours, covers 28 of the 39. Ten songs disagree with their page under all three trackers while the trackers agree with each other within 4 percent (among them a page saying 75 where the trackers say 106, and 124 swing against 164); they are listed in the brief, not re-graded.

Predictions and Disclosures

Mostly wrong in the details, right in the shape: the trackers landed four points under the registered 80 percent strict on straight 4/4 and two under the 95 folded line; the small checkpoint beat the large one on swing meter by eight points where five was the registered margin; the instrumental break rate was 6.5 percent against a predicted 5. Right: madmom within eight points of Beat This on 4/4 and worse on 6/8, the instrumental fix rate under 25 percent, the fold direction, and the interval width. One labelling disclosure: the fold-direction prediction was written in tempo terms (a 240 song read at 120 called the half fold) while the grader's ratio is on the bar, where that same reading is a two; the prediction is scored under its own words and the brief says so. The twelve Study 34 and 35 songs sit inside this corpus and move no rate. The run took five hours of CPU on this VM, not the fourteen the registration estimated, and cost nothing in model spend.