Study 17 · Listening
The transcription successor: the neural loser wins the real-audio rematch, and the reconciler meets its first breaks
Listening, the rematch
Study Overview
The Question Study 13 Left
Study 13's closing guidance was explicit: if the chord feature matters, the next investment is Basic Pitch's successor on real timbres, not more model. Study 17 measured three zero-model successors, all built from things production already holds: per-stem transcription (the app stores the stems separately, and Study 13's path summed them first, feeding Basic Pitch the polyphonic mixture it fails on), bar-pooled chroma into production's own template matcher (nothing requires its 12-bin profile to come from transcribed notes), and ChordMini, the neural reader that lost Study 10's synthetic chord gate and had never been scored on real audio with bar-aligned chord truth, because that truth did not exist until Study 12. A conditional reconciler arm rode on the winner. The anchor gate passed exactly: the current path reproduced Study 13 arm 1 to the digit, 29 of 76 gated and 24 of 86 exploratory, before any new number was read.
The Rematch
ChordMini on the mix, its native domain, reads the gated recording at 59 of 76 bars, 77.6 percent, and the exploratory song at 53 of 86. That passes the successor gate at 37 with twenty-two bars to spare and, more consequentially, passes 44 of 76, production's reconciled number from Study 14: a zero-model front end now beats code plus model, which flips the ship order. Study 10's synthetic verdict reversed on real audio, the home-field lesson's third instrument class after sections (LinkSeg) and the same-piece constant, and the rule is now three for three: no detector decision for real recordings gets made on synthetic numbers. A disclosed secondary read on the summed harmonic audio beat the mix on the gated song and collapsed on the exploratory one, so the registered choice of the mix survives its own alternative.
Roots, Not Dressings
The gated chart's truth is 74 power-chord and 50 sus4 slots in 76 bars: this band writes roots and colors, not sevenths. ChordMini's gated credit came entirely through the registered power-chord allowance, with zero strict-first matches, because its vocabulary is root-level triads; it will never spell F#5 or Esus4, so the draft layer decides the dressing. A disclosed root-only read makes the real ranking visible: the current path holds 53 of 76 roots, per-stem 65, chroma 68, ChordMini 70, and 75 of 84 on the exploratory song. Every successor out-reads the anchor at the root level; the registered differences are quality dressing. The sharpest case is the chroma arm: 0 of 76 under the registered grade while holding 68 of 76 roots, because its pooled mass dresses nearly every right root in a hallucinated seventh, the exact over-read class its pre-registration self-proof surfaced and the brief disclosed, and the allowance admits only major, minor and power qualities. A quality fix needs a tuned mass threshold, which the arm was registered not to have, so it belongs to a registered follow-up or nowhere. Per-stem transcription split by stem inventory, and the mechanism is legible: letter-lands has no vocal stem and improved to 32; eat-all-the-clocks includes one, whose transcribed melody pollutes the merged harmony, and fell to 15 of 86. One song each way is a hypothesis, not a finding, but it ships with a rule attached: if per-stem transcription goes anywhere, vocal-labeled stems stay out of the chord evidence.
The Reconciler's First Breaks
Arm E stacked Study 14's measured production reconciler path on the winner, with the registered design choice disclosed in the brief: ChordMini produces no note events, so the model saw ChordMini's chords beside the current path's transcribed note lines. The result is the study's most instructive number: 59 fell to 47, four fixes against sixteen breaks, the first breaks in three studies of reconciler runs, and the double-fix gate registered against Study 9's agreement trap failed for the first time, exactly as built. The mechanism is clean: ChordMini's chords are better than the transcribed notes support, so a pass that corrects the chords toward the notes destroys information. On the exploratory song it changed nothing. The reconciler's measured value is conditional on its evidence channel being the one that produced the draft, and the verdict delivered to the app says so: keep it on the Basic Pitch path it was measured on, and never run it on a non-note front end.
Ship Guidance, and the Scorecard
Delivered to Stem Shovel in order of weight: ChordMini on the mix is the measured chord front end for real recordings, about 30 MB of hash-pinned ONNX weights already vendored since Study 10, four to six seconds a song on CPU, zero API spend, lifting the real number from 38 to 78 percent registered and from 53 to 70 of 76 at the root level; its root-level vocabulary ships with it, so the chart-draft layer renders the dressing, and for this catalog root-plus-power spelling covers most of the page; and the reconciler stays off its output. The scorecard kept the program honest: two predictions right, one half, three wrong, with the chroma favorite scoring zero under the registered grade and the reconciler prediction wrong in the instructive direction, two or fewer breaks predicted against sixteen measured. About $0.30 of the $2 fence, all of it in the two reconciler calls; arms A through D cost nothing.