The stereo file needs no stems: the mono downmix drafts above the stems sum, and every clever split loses

The chart-draft pipeline · the stereo file

Study Overview

The Question, and What Was Already Settled

Kevin asked whether the auto-draft works on a stereo file, and whether the file is worth treating as two stems. The grid half was already answered: every study since 28 ran on mono downmixes of these exact stereo mixes (the fetch converter runs ffmpeg -ac 1), so a fresh stereo upload gets the validated grid for free. The untested half was the draft, whose Study 31c reference numbers came from transcribing the non-drum stem sum. A zero-dollar scoping measurement set expectations: the four mixes correlate L/R at 0.83 to 0.90, the side channel sits 10 to 13 dB under the mid, weakest below 200 Hz. Kick and bass live dead center; each channel is a near copy of the whole band, not anything stem-like.

Five Arms, One Variable

Per song, five transcription inputs went through one identical pipeline: the non-drum stem sum (R, the ceiling), the plain mono mix (A, the naive path), L and R as two pseudo-stems (B, the question as asked), mid and side (C, the steelman), and the HPSS harmonic of the mono mix (D, the DSP comparator). The grid, ship vote, placement and nchord hints were computed once on the mono mix and shared by every arm, so the transcription input is the only variable. Grades are root scores against the chart. The pilot ran letter-lands under a $0.30 fence; the registered matrix, peaceful-dreams alone under $0.60, ran on Kevin's go after the pilot's results were committed.

The Mono Mix Beats the Stems

The reproduction gate passed to the bar twice: arm R landed exactly on the frozen references, 72 of 76 on letter and 151 of 153 on peaceful. Then the unregistered headline, twice: the plain mono mix drafted ABOVE the stems sum on both songs, 74 of 76 (97.4 percent) and 152 of 153 (99.3 percent). The registered prediction said drum bleed would cost bars; it was wrong in direction. Basic Pitch transcribes drums weakly, the mix yielded fewer notes than the stems sum on letter, and the draft model reads through what bleed remains. S33-G2 passed on both songs and closes Kevin's question: the L/R split went one bar under mono on letter and tied it on peaceful, at double the transcription cost both times. Not worth it.

Mid/Side Was Noise, HPSS Is a Trap

Mid/side was the pilot's best arm at 75 of 76, one bar above mono and inside the disclosed sampling noise; the matrix flipped the sign hard, 146 of 153, six bars under mono. The side channel added about 2,270 cymbal, reverb and width notes on peaceful, and enough of them read as harmony to poison six bars. HPSS failed its gate on the pilot (two bars under mono, published as found), then tied mono on the matrix song: inconsistent per song and never better, pure cost. The zero-dollar panned-synthetic sweep names the mechanism: with centered drums and the chord instrument panned at width 0, 0.5 and 1.0, HPSS sits at the 99.6 to 99.7 percent chord-tone purity ceiling at every width, and the side channel is pure harmony exactly when the mix pans harmony. Clean synthetic sustains are the filter's best case; real strummed attacks and vibrato read as percussive, and the filter takes harmony with them.

The Production Answer, and Two Disclosures

A chartless stereo upload goes into the pipeline as its mono downmix: same grid, same hints, same draft, stems-ceiling quality on both songs measured. No channel splitting, no mid/side, no source separation. Two disclosures rode with the results: the matrix runner as first committed would have re-run the pilot song's five drafts outside the registered scope and was corrected before any paid call; and the matrix fence was exceeded, $0.68 against the registered $0.60, because the two-buffer arms carry nearly twice the notes and peaceful's drafts grew to about 38k tokens in each. The study total is about $0.98 across 1,158 cells.