The Bar That Was Three Bars Long
The last chapter ended two studies deep in failure. The auto-chart pipeline misread a century of recordings, and handing it the true tempo did not save the old ones: the grid drifts on unevenly cut 78s, and the note transcriber starves on acoustic-era audio even where the grid lands to the bar. Both walls got hold-for-review rules in Stem Shovel the same day, read off those same fifteen recordings, which made them hypotheses in need of material they were not tuned on. And buried under all that failure sat an untested claim of good news: the two modern recordings in the corpus had nearly passed. So the app wrote the fairest brief in the program so far: run declared mode, the exact shape of the planned public tool, on the material that tool will actually meet. Modern recordings, made this century, by someone who tells you the tempo. Score the new holds in advance, registered before the run, so that whatever they do out of sample is a finding and not an argument.
stem-bench built the corpus from the catalog of Josh Woodward, an artist who publishes his songs under Creative Commons and, better, publishes beside each one the thing this study needed most: the tempo, the key, the meter and the full chord chart, in his own hand. Ten songs went in, from a sparse banjo ballad to a full-band rocker to a strummer whose page declares 240 beats per minute, including both compound-meter songs the catalog's charts could support, one swung, one straight. Three songs proposed first had to be swapped before registration, disclosed in it: their pages carry only guitar tablature or no chord symbols at all, and a registered truth table cannot be built from shapes the bench would have to interpret.
Truth from the Artist's Own Hand
A chart written for guitar hides a trap: the capo. The artist writes the shapes he plays, and the recording sounds somewhere else entirely, five or seven semitones up. The bench does not get to eyeball that shift, so each song's transposition was decided by a gated calibration before registration: slide the written chart against the recording's chroma at all twelve offsets and accept the winner only when it beats the runner-up by a clear margin. Five songs passed the gate, and here is the part that makes the method worth the trouble: every accepted value equals the capo the artist states on his own page. Capo five, capo seven, capo two, no capo, capo five, confirmed blind, five times, on ground truth the instrument had never seen. The three songs whose margins were too thin to trust got graded transposition-invariantly instead, on the intervals between successive chords, which no capo can disturb.
The corpus freeze threw in a free headline before a single model token was spent: under the declared tempo from each artist page, every one of the eight new grids locked within 1.4 percent, most within a rounding error. The onset comb that could not hold a 1917 lathe holds modern material exactly as well as the program had been hoping since Study 34.
The Texture Flips
Five of ten drafts graded good. Three of them carried perfect root sequences, every chord change in the artist's chart recovered in order, and one of those three is the song declared at 240 beats per minute. Three more drafts sat at 83 to 86 percent, just under the registered bar of 90. The transcription that starved on acoustic-era audio fed these drafts eleven to fifty notes a bar. The registered gates still fail by letter, five good of a required seven, and this series reports that plainly. But for the first time in three studies the honest summary sentence is positive: on the material a public upload tool would actually meet, declared mode drafts a usable chart most of the time.
Most of the time. Two songs landed in the worst class the program has, confidently wrong and saved, and they are not a random two. They are both the 6/8 songs.
The Bar That Was Three Bars Long
The failure signature was strange on its face. Both compound-meter drafts scored near-perfect on chord qualities, the part of grading that ignores pitch entirely, while scoring around 60 percent on roots. A draft that knows minor from major from seventh but misses the notes is not hearing the music wrong. It is standing in the wrong place.
The cause is a units mismatch, and it is mechanical all the way down. A musician stating the tempo of a 6/8 song gives the dotted quarter, the pulse you would tap your foot to, two to a bar. The product's declared path reads a typed tempo as beats of the bar per minute, six to a bar in 6/8. Same number, different unit, and the declared bar comes out exactly three times too long. The pipeline then did everything right downstream of the mistake: it found a grid within 1.4 percent of the bar length it was asked for, because a span of three real bars repeats just as periodically as one, and it drafted one chord per seven-second bar over music that changes harmony every two seconds.
Here is the uncomfortable part for the new safety holds, the reason this study registered them in advance. The grid-drift hold measures the landed grid against the declared value. When the declared value itself is misread, the hold compares the mistake to the mistake and reports everything fine, zero drift, chart saved. No threshold can fix a units problem, because the problem is upstream of every number the threshold sees. The app had fixed this exact disease on a different code path the very morning of the run, where a chart file's own tempo line was being read in the wrong unit. The chartless path needs the same cure, an input adapter that asks, or infers, which note a visitor's tempo counts, and the bench's follow-up offer is already on the table: two drafts, about twenty cents, to re-run both songs under the registered gates once it ships.
The Refusal That Kept Only Mistakes
The other registered hypothesis did not survive contact either, and its failure is the loud kind. Study 34b had found, exploratorily, that the detector's tempo confidence separated its wrong detections from its right ones at a threshold of 0.84, with the comforting property that its rare errors flagged right answers as doubtful, the safe direction. Out of sample, on the modern corpus, the threshold inverted. The detector got the tempo wrong on eight of ten songs, and the confidence refusal fired on eight of ten, which sounds fine until you look at which eight: it refused both of the detections that were right, and the only two detections it let through were wrong. A gate that was supposed to catch mistakes caught everything except mistakes. The registered prediction said it would fire on every wrong detection and at most two right ones; the data falsified it in the worst available direction, and the bench's answer to the app is one sentence long: do not ship tempo confidence as a refusal signal.
The Grid That Finally Followed
One exploratory arm, registered without a gate, reached back to the recordings this chapter was supposed to leave behind. The two 1917 jazz 78s fail every fixed grid because the sides were cut unevenly: no single bar length exists, so a comb that assumes one drifts off by seven to nine percent no matter what it is told. The arm replaced the comb with a tempo-following beat tracker, the standard dynamic-programming kind, seeded at the declared tempo and allowed to breathe with the record. On Livery Stable Blues it tracked 510 beats into 128 bars against an annotated 125, and the draft on that grid scored 112 of 125 roots, just shy of 90 percent, on a recording where every fixed-grid attempt across three studies had scored 17 to 31. One run proves nothing alone, which is why this arm carried no gate, but it answers its registered question cleanly: a tempo-following grid earns a design. The other 78 stayed at the floor with its bars landing, which is the starved-transcription wall again, and no grid will fix that one.
The Ledger
The whole study cost one dollar, exactly, eleven paid drafts against a three dollar fence. The pre-registration census is disclosed where it fell short: the brief wished for two songs with real pickup beats and the corpus, measured, carries none, so that bound's catch rate stays unmeasured for another study. Nine of the ten modern songs are one artist's voice and production style, a limit the registration states rather than hides. And one grading defect surfaced after the run and was corrected the same day, before any of these numbers were published: the first pass compared the drafts' sounding chords against the charts' written shapes without applying the calibrated capo, which no earlier corpus could have exposed because no earlier corpus had a nonzero capo that passed the gate. The tell was the same signature as the units bug, wrongness confined to exactly the capo songs while their pitch-free quality scores sat near perfect. The grader gained a transposition-aware scorer with unit tests, the nine stored drafts were regraded without a single new model call, and both gradings, before and after, sit side by side in the committed results for anyone to check.
The registrations, the truth instruments, the raw drafts and the full report live in the stem-bench repository; the entry is Study 35 on the research dashboard. The music is Josh Woodward's, shared under Creative Commons, down to the chord charts that became this study's truth. The pipeline that failed a century of recordings mostly passed its own century, and the two ways it still failed were not music problems at all. One was a unit. The other was a threshold that learned the wrong lesson from the previous study, which is exactly the kind of lesson you can only unlearn out of sample, in a registered run, where the prediction was written down first.
