The Chart Nobody Typed

The last chapter closed on three findings. The 6/8 failures were a units bug: a compound song's declared tempo was read as beats of the bar, making the bar three times too long. The tempo-confidence refusal inverted out of sample and let through only mistakes. And a tempo-following beat tracker, seeded at the declared tempo, recovered Livery Stable Blues, a 1917 78 that every fixed grid had failed. That raised the question for the planned public tool in Stem Shovel: declared mode asks a visitor to type the tempo and pick the meter, so what if the visitor types nothing and a published beat tracker supplies the bar? Ten studies over two days answered it, mostly in the tracker's favour. Two walls still stand: free tempo, and knowing which songs are in 6/8 at all.
The Field's Trackers Find the Bar
Study 36 put the field's beat and downbeat trackers against the production detector on the 23 recordings with bar-level truth, at zero model spend. Two families competed: Beat This!, with a large checkpoint (final0) and a small one meant for the browser (small0), and madmom, whose dynamic Bayesian network (DBN) models the bar. On the ten modern Josh Woodward songs, scored as bar strict (within 4 percent) / bar right at some metrical fold / meter class right, production scored 3 / 7 / 8. Beat This final0 scored 6 / 9 / 8, small0 7 / 8 / 8, madmom 7 / 10 / 8. The registered decision rule asked for 7, 10 and 8, and only madmom met it, so madmom's bar carried the first automatic drafts. P1, that Beat This would clear that bar, failed on all three clauses; P2, P3 and P5 passed; P4 failed as a conjunction, partly because every Beat This variant put its downbeats on beat 3 of Dippermouth Blues' bars.
On the 78s, every neural tracker read Livery Stable Blues' bar at 1.46 seconds where production read 2.71, and madmom laid 126 bars against 125 annotated, with a downbeat F-measure of 0.96 at 70 milliseconds. Dippermouth came out 114 of 114 bars, St. Louis Blues 52 of 53. That is the last chapter's result with no declared tempo. And production's automatic mode gets the 6/8 song gravity's bar almost exactly (2.466 seconds against 2.5), so Study 35's failure there was the units bug alone. Beat This has no compound class and reads both 6/8 songs as three-beat bars.
Two Hundred Sixty-One Songs
Study 36b took the whole catalog: 261 recordings from the artist's catalog bundle, purchased for the bench, each with the tempo, meter and feel he states on the song's page, 244 of them paired with an instrumental mix. Production's bar was strict on 36 percent and folded on 67; Beat This final0 reached 70 and 90, small0 69 and 90, madmom 66 and 89. On straight 4/4 the trackers were strict on 73 to 76 percent against production's 38. On 6/8 nobody was good: strict 8 to 35 percent. Instrumental mixes fixed 10 of Beat This's 74 failures and broke 11 of its 170 passes: the voice is not the problem.
Q2 asked for 80 percent strict on straight 4/4 and got 76; neither decision rule fired. Of final0's 39 folded-but-not-strict 4/4 readings, 28 were a bar half the declared length, on songs declared at a median 72 bpm. The trackers live in the 80 to 160 lane, and a half bar is a finer grid that a chart should be able to merge back, a claim Study 37 went on to test.
The Detector's Own Worst Call
36b also found production's largest error class, and it was the detector's own compound-meter call. It named 6/8 on 63 of the 223 songs in 4/4, every one through a subdivision test reading 3, and 59 of those bars fold to nothing. When it named 6/8 on a song that really is in 6/8, 21 of 37 times, the bar was strict on 2. Its tempo confidence did not separate right from wrong either (medians 0.762 against 0.685), so the advice not to ship the refusal held across 261 songs.
Study 40 swept the app's candidate fixes over the catalog through a probe pinned to the detector on 261 of 261 songs. No witness separates true 6/8 from false. The duple-bar witness at r 1.00 drops the true calls from 21 to 4 while 10 false ones remain. But the call has no bar value to protect, so gating or removing it raises production's strict bar from 95 songs to 112 to 114, folded from 176 to 200 or 201, meter right from 180 to 214 to 217, with zero to two songs made wrong. Taking the finer bar when the detector's double-tempo candidate is strong is net negative at every strength (2 fixed and 17 broken at the strictest), and no tempo cap rescues it. P2 and P3 passed; P1 and P4 failed, and P5 missed by two points.
The Half Bar Costs Nothing
Study 37 asked whether the octave gap needs closing. Part C tried a witness that picks the bar on which chord changes land. In 4/4 it was net negative on every base: on final0 it fixed 1 of 34 folded readings and broke 30 of 165 strict ones, taking the strict rate from 74 to 61 percent. The chord reader emits 60 to 124 changes a song, and of the witness's 77 moves, 26 went to the half bar and 11 to the double. It helped only on 6/8, for madmom and production. Six of seven predictions failed; only C5, the abstention rate, held.
Part D bought five drafts. Three songs drafted at half the declared bar scored 20 of 21, 16 of 18 and 33 of 33 roots against 18, 15 and 28 at the declared bar, and two crossed into the good class. The two 6/8 songs, drafted at the bar the corrected tempo input would hand over: gravity went from 16 of 26 to 23 of 26, my-favorite-regret from 26 of 42 to 40 of 42, both out of the confidently-wrong class they had held in Study 35 and one of them good. D2, D3 and D4 passed. D1 failed under its own words, because two drafts improved by more than its two-root band.
Nothing Typed
Study 38 drafted the first charts with no input at all: madmom's downbeats as bar lines, its median beat as tempo, its beats per bar as meter, the detector's key. All ten modern songs landed within two roots of the declared-grid draft or above, eight good, none confidently wrong. Gravity scored 24 of 26 and my-favorite-regret 40 of 42 with no meter typed. A1, A3 and A4 passed and A2 was half right. Decision rule S38-D1 fired: automatic mode is the tracker's bar as read, shorter bar preferred, with the Study 35 confidence refusal retired.
On 36 solo acoustic guitar excerpts from GuitarSet, with each player's lead sheet as truth, the annotated bars gave 99 percent mean roots and madmom's bars 96, a paired cost of 2.9 points. Without drums, Beat This placed the better bar lines (strict 86 percent against 61, downbeat F 0.97 against 0.69). On acoustic material the next target is the chord reader, not the bar.
Where It Still Fails
Study 38's third arm was free tempo: 24 Winterreise songs for voice and piano, the only free-tempo material with chord truth the search found, a proxy for unclicked demos. Every draft was confidently wrong, with a roots median of 41 percent. The music changes chord 1.5 times per tracker bar, and madmom's downbeat F was 0.39 where Beat This's was 0.68: C4 had predicted the reverse.
Study 39 ran eight systems over 48 Winterreise performances. Beat This final0 had the recall (a median 0.86 at 70 milliseconds) but put a half bar on 14 of 48. Every DBN held a bar length and lost the beats. Loosening madmom's tempo penalty bought 0.08 of recall, half the predicted 0.15. None was chartable on more than 11 of 48 against a floor of 36, so no free-tempo tracker was adopted. P1, P3 and P4 passed; P2 and P5 failed.
Study 43 tested the chart format: two cells a bar instead of one, on Beat This small0's bars. On the lieder the roots median rose from 0.36 to 0.58 and every song stayed confidently wrong. On the guitar the premise was wrong: the dense lead sheet behind Study 38's three confidently-wrong guitar drafts is one chord a bar with repeats, and on small0's bars four of six takes already scored 14 of 14 roots. What holds five of six below good is chord quality. Unlooked-for, D911-19 drafted good (37 of 41 roots) on small0's bars as read, the first Winterreise song to reach good. The half grid lifted 7 of 20 items and lowered 2, so no grid change was recommended. P1 and P2 failed; P3, which predicted a lieder median of at least 0.45 with no song good, and P4 passed.
One Tracker, Checked to the Frame
The app wanted to ship the small checkpoint alone. Study 41 redrafted the ten modern songs on Beat This small0's bars: within two roots of madmom on 10 of 10, eight identical, two better, none worse, none confidently wrong. The app's rule passed (P1, P4 and P5 passed; P2, which expected one failure, and P6 failed; P3 went unscored). It also caught a label blind spot. On gravity small0's most common bar holds six beats, and the old rule labelled it "4/4 at 143." Labelling from the most common count, with six read as 6/8 at the tempo divided by three, lands on 48, the declared tempo. Of the app's three draft holds, the drift hold came out (it held 17 of 46 drafts, 12 of them good). And the bench's own browser-sized export from Study 10 carried the right weights at the wrong window: its attention breaks past about 300 frames, and at 5 seconds of context it is a different tracker, with bar counts off by up to 82. Study 10's N-beat numbers came from that tracker.
Study 42 traced the cause to a rotary position embedding cache baked in at the traced length, and made two exports matching the package's frame outputs to within 7e-5. The static one reproduced the package's beats and downbeats exactly on all ten songs. In WebAssembly, single-threaded under Node, it takes 4.5 seconds per 30-second chunk, so the bench recommended the browser with the server as fallback. P1, P2, P3 and P5 passed. P4 failed: at default threads ONNX Runtime took 1.5 times as long as torch.
Study 44 measured two rules the app added itself. Joining adjacent half bars moved the census bar count toward the artist's on 73 songs and away on 61, but it rescued gravity, 118 bars down to 94 against 94 declared. Four drafts on the joined grid scored identically, so joining is free where it fires. The steadiness rule enriches for a wrong bar without separating. At the app's 0.6 it removed 35 census readings, 22 of them right, and among the 22 drafts on small0's bars it removed two sound ones and missed the one confidently wrong. The registered criterion landed on 0.7. The bench's labelled judgment is 0.4, the least that keeps mop heart, a 3/4 song at 224 bpm that the big checkpoint misreads too, off the chart. P2, P3, P5 and P6 passed; P1 and P4 failed. The app's stored beats and downbeats matched the package's to the frame: 1,772 beats and 478 downbeats.
The Ledger
Studies 36, 36b, 39, 40 and 42 cost nothing; Study 37 spent $0.452 of a $0.60 fence; Study 38 spent $3.84 of $8, plus $0.07 on an unregistered retry; Study 43 spent $1.54 of $3; Study 44's joined-grid arm spent $0.35 of $0.80. Study 41 spent $2.01 against a $2.00 fence; the fence is checked before each draft, and the last draft's own cost carried it over by a cent.
Two instruments were corrected on the record. Study 37's first pass measured the bar from the mean beat interval and failed to reproduce Study 36b; it is preserved, and the fix was committed with the results, not before the re-run as the program's hard rule asks. Study 40's first two feature extractions failed the pin to the detector and were set aside. Six drafts stopped at the gateway's content filter (three in Study 38, even after a retry, and three in Study 43); they are labelled refused and were not worked around. And checking this chapter against the record turned up errors in these studies' write-ups, in Studies 36b, 37, 40, 41 and 43: four sentences whose numbers disagreed with their own tables, and a prediction in Study 43 scored against a threshold its registration does not contain (it held); all were corrected in place on 2026-11-13 with the first wording kept, and none changed a decision. One Winterreise recording's license, read after download, is non-commercial and no-derivatives; it is used for measurement only, and a public-domain 1933 recording was added so the free-tempo bar result rests on rights-clear audio too.
The registrations, tracker readings, drafts and full report live in the stem-bench repository. The entries are Study 36, 36b, 37, 38, 39, 40, 41, 42, 43 and 44 on the research dashboard. On modern songs, a visitor who types nothing now gets a chart within two roots of, and often better than, one drafted from the artist's own numbers. What remains is music that does not hold still, and songs whose meter nothing here can yet name.