The File That Needed No Stems
Stem Shovel can now draft a chord chart for a fresh song on its own: tempo measured, bar one voted by the witnesses the last chapter described, and the chords read from a machine transcription of the song's stems. That last word is the catch. The pipeline was built for stems, the isolated instrument tracks a band uploads from a session. Most of the world's recordings are not stems. They are one finished stereo file. So the question that opened this study was asked plainly: how well does the auto-draft work on a stereo file, and is it worth treating the file as two stems, left and right, and running it through the same process?
stem-bench, our open, pre-registered benchmark series, started where it always starts, with a measurement that costs nothing. The four corpus recordings, Badverbs' own, correlate left to right at 0.83 to 0.90. The side channel, what distinguishes left from right, sits 10 to 13 dB under the mid, and it is weakest below 200 Hz, because kick and bass live dead center in these mixes, as they do in most. That one measurement reframed the question before a dollar was spent. A left channel is not half the band. It is a near copy of the whole band, slightly re-weighted. Splitting a stereo file into L and R does not give the pipeline two stems. It gives it the same mix twice.
Five Arms, One Variable
The registered design held everything fixed except the one thing in question. Per song, the beat grid, the bar-one vote, the placement and the chord hints were computed once on the mono mix and shared, so five arms differed only in what the transcriber heard. The reference arm transcribed the non-drum stem sum, the thing the pipeline was built for. The naive arm transcribed the plain mono downmix. One arm split the file into left and right pseudo-stems, the question as asked. One split it into mid and side, the steelman, since the side channel at least isolates something. And one ran harmonic/percussive separation on the mono mix and transcribed the harmonic half, the signal-processing stand-in for "remove the drums" that needs no stems at all. Every arm's draft was graded mechanically against the band's chart, bar by bar.
The reproduction gate passed to the bar, twice: the stems arm landed exactly on its frozen reference from two studies ago, 72 of 76 bars on Letter Lands and 151 of 153 on Peaceful Dreams. The ruler, for once in this series, needed no correction.
The Mono Mix Beat the Stems
Then the headline nobody registered. The plain mono downmix, the one arm that does nothing clever at all, drafted above the stems sum on both songs: 74 of 76 against the stems' 72, and 152 of 153 against the stems' 151. The registered prediction said drum bleed would cost bars. It was wrong in direction. The transcriber hears drums poorly to begin with, the full mix actually yielded fewer notes than the stems sum on one song because the band masks the kit, and the drafting model reads through what bleed remains. A stereo file does not approach the stems ceiling. On both songs measured, it is the ceiling.
The question that commissioned the study closed just as cleanly, in the registered direction, twice. Left and right as two pseudo-stems scored one bar under plain mono on the first song and tied it on the second, at double the transcription cost both times. The gate that encoded "not worth it" passed on both songs. The scoping measurement had said why in advance: each channel is the whole band again, so the second transcription adds duplicates, not information.
The Clever Options Lose Harder
The pilot left one loose thread, and it was the interesting kind. Mid/side, the steelman split, was the pilot's best arm, one bar above mono, inside the sampling noise the registrations have disclosed since the gateway stopped honoring temperature. Was the edge real? That is precisely what the registered matrix run existed to decide, and it decided with prejudice: on the second song mid/side crashed to 146 of 153, six bars under mono. The side channel added about 2,270 extra notes of cymbals, reverb tails and stereo width, and enough of them read as harmony to poison six bars. The pilot's one-bar edge was noise with the sign flipped.
The separation arm told the best story of the run. On the real songs it failed its gate on the pilot, two bars under mono, then tied mono on the matrix song: inconsistent per song, never better, pure cost. But the study also carried a zero-dollar synthetic sweep, songs rendered with centered drums and the chord instrument panned at three widths, scored by chord-tone purity. On those clean synthetics the separation is essentially perfect, 99.6 to 99.7 percent purity at every pan width, right at the ceiling. That contrast names the trap precisely. Median-filter separation is flawless on steady synthetic sustains, and real music is not steady: strummed attacks and vibrato read as percussive, and the filter takes harmony with them. The sweep also confirmed the side channel is pure harmony exactly when a mix pans its harmony, which these mixes, like most modern mixes, do not.
The Ledger
This series publishes its own misses, and this study has two. The matrix runner as first committed would have re-run the pilot song's paid drafts alongside the new one, outside the registered scope; the defect was caught and corrected before any paid call, with the fix disclosed in the commit. And the matrix fence was exceeded: $0.68 spent against the registered $0.60, because the two-buffer arms carry nearly twice the notes and the second song's drafts grew to about 38,000 tokens each. The estimate was built on single-buffer prompts and did not survive contact with the steelman. Found after the run, published as spent. The whole study, both songs, ten drafts, twenty transcriptions and the sweep, cost about 98 cents.
The production answer is the simplest sentence the pipeline has earned yet: a chartless stereo upload goes straight in as its mono downmix. Same grid, same hints, same draft, at or above stems quality on both songs measured. No channel splitting, no mid/side, no source separation. Every clever option lost to the downmix, and the one that looked like it might win was the one a second song unmasked. The registrations, the gates, the fence ledger and the full report live in the stem-bench repository; the entry is Study 33 on the research dashboard, beside the grid vote and adversarial corpus that made a chartless draft worth testing in the first place. The app is Stem Shovel, the recordings are Badverbs' own, and the file, it turns out, was enough.
