The witness vote finds every downbeat, exposes the bench's fourth ruler error on the way, and carries the chartless draft to chart parity

The chart-draft pipeline · the witness vote

Study Overview

Seven Witnesses, One Class Each

Every candidate is a whole-beat shift of the Study 28 anchor, so each witness votes for a beat class or abstains: time zero (weight 2), the last strong onset (1, abstaining on fades and ritardandos), the beat count between the ends (1), harmonic rhythm from per-beat chord labels (2, run on the bench's own reader as a disclosed substitution for production's chord runs), crashes with beat-long decays (1), the drum and bass stems (2, bass can veto), and checkerboard section boundaries (1). The weighted vote moves off the base only when the winner leads it by two points and the runner-up by one. Weights, margins and gates were committed before the run and never tuned.

The Failure That Was a Finding, and Correction 4

Study 31 as registered failed all three gates, and the pattern said why: three independent witnesses agreed with the truth exactly on three songs and agreed with each other on eat, one beat before the registered truth. The adjudication confirmed the witnesses. Eat's +0.415-second anchor was Study 25's hint-fit, 55 of 84, made against the phantom-bars chart before Correction 2 and never re-derived after it, and the hint instrument is measured flat across a two-beat range on this song, 65, 66 and 66 of 69, so it never had the power to pin the phase. Four independent constructions put bar one at minus 0.06 to minus 0.07 seconds: the chart-informed DTW, whose constant 0.474-second offset in Studies 27 and 29 had been misread as eat's quarter-bar edge; the chord-change histogram at margin 35; all twenty section boundaries; and time zero. v5 was right on eat all along, to a few hundredths of a second. The corrections ripple in place: Study 28's comb phase missed all four downbeats, not three, and Study 30's scorers re-read by class go two and three of four rather than zero and two.

The Vote, the Placement Rule, and Parity

On the corrected ruler, Study 31b's vote found the true class on all four songs with none undecided, and harmonic rhythm alone went four for four, the witness the app's brief says it ships first; time zero went three of four, abstaining on mop and never wrong. The post-mortem is honest: crashes voted one of four, the stem witness vetoed itself on every song, and the two ends witnesses abstained everywhere. The re-draft gate then exposed the family's second instrument defect: a beat class names a position modulo one bar, and the runner had placed voted anchors a whole bar from the song's bar one, collapsing the drafts below even the backbeat base and retro-explaining Study 30's worst number. Study 31c registered the placement rule, bar one contains the first sound, and re-ran the arm: letter 71 of 75, exactly its chart-grid reference, and peaceful 152 of 153, above its reference. Chart parity without the chart. Family spend 39 cents of a 50-cent fence; the gateway ignores temperature on this model family, so draft scores carry a few bars of sampling noise, disclosed in the registration. The condition that stands: the humility screen was vacuous once the ruler was fixed, so the vote ships behind the app's fifth and sixth songs.

The App's Review, and Study 31d

The collaboration ran both directions: the app's agent read the committed results and found what the bench had missed. The onset envelope's extractor caps at 90 seconds, so the last-onset, both-ends and stem witnesses had never seen the back half of letter (138 seconds) or peaceful (254 seconds); their abstentions were unmeasured, not measured. And time zero's four committed distances, 0.079, 0.058, 0.068 and 0.070 seconds at four different tempos, are one number: the envelope's documented lead, measured by a witness that does not read the envelope. Study 31d registered both fixes and ran for free. Uncapped, the ends witnesses are honestly dead on this corpus, abstaining by ritardando on all four full songs, and the stem witness still vetoes itself. The lead-corrected time zero is four of four with distances of 0.02, 0.03, 0.008 and 0.000 beat, mop's earlier abstention revealed as the fence catching the lead. Both vote variants pass the gate. The final roster the app ships: lead-corrected time zero, harmonic rhythm and section scale, each four of four alone; the vote over them; undecided keeps the base; the tap stays out. Every truth in the family is the downbeat at or near time zero read 0.07 seconds early, a uniform disclosed ruler bias that beat-class scoring is immune to.