The adversarial corpus gives the humility gate teeth, the ship vote survives every honest fooler, and the one wrong answer names the rule the app needed

The chart-draft pipeline · the adversarial corpus

Study Overview

The App Wrote the Brief

With no fifth and sixth songs available, the app's agent drafted the adversarial study itself: renders where the base is right and no witness may move it, starts that are not a bar line (a one-beat pickup, a half bar of silence, two seconds of silence, a four-click count-in), harmonic rhythms built to fool the chord witness (changes on one and three equally, one chord per two bars, a one-chord vamp that demands abstention, anticipations an eighth early), a backbeat-heavy mix, 3/4 and 6/8 beside 4/4, and the four real recordings re-cut by 0.3, 1.0 and 2.5-beat pads plus a one-beat trim. The bench adopted the brief whole and registered six reconciliations before the run, including its own mini-synth, since the fixture synthesizer is bar-granular and cannot place an anticipation. The ship vote under test is exactly what the app shipped to its dev branch the same day: lead-corrected time zero at weight two, chord changes under a 1.2 confidence ratio at weight two, section scale at one, the 31b humility rule, undecided keeps the declared grid.

The Bench's Kick Could Not Be Heard

Run one failed the humility gate before any witness got the chance to: every synthetic correct base measured one beat late, because the mini-synth's kick was a pure 65 hertz sine and the spectral flux onset envelope barely registers it. The renders were backbeat-dominant in flux at any kick amplitude, the comb landed on beat 2, and the construction never produced its premise. That is the program's recurring lesson wearing a new costume: the backbeat does not only fool detectors, it fooled the bench's own synthesizer design. The disclosed fix in Study 32b is six milliseconds of broadband beater click, committed with a test that measures the flux difference. Meanwhile the real songs had supplied correct bases on their own, and the gate found its first genuine violation there: the crash witness voted to move letter-lands' correct base on all three of its correct-base cuts. It was already outside the ship vote; it is now formally dead. No shipping witness ever moved a correct base in either run.

No False Moves, Both Directions, One Named Rule

With the fixture honest, the re-run passed the no-false-moves gate outright: eleven adversarial synthetics, zero wrong decisions, one honest undecided, and that undecided was the two-chord-bar fooler that had produced run one's confident wrong vote. The instructive detail is which mechanism held: the registered prediction said the chord witness's 1.2 ratio rule would abstain on two equal change classes, but the histogram favored one class past the threshold and voted wrong anyway, and the HUMILITY MARGIN refused the move. The count-in was self-defeating in 4/4 as predicted, four clicks being exactly one bar. The anticipation fooler failed to fool: the chord model snaps eighth-early changes to their bars. The backbeat mix got both directions across the two runs, a wrong base moved to the truth five to nothing in run one, a correct base kept five to nothing in run two. On the real re-cuts the vote went fifteen right, four undecided, one wrong, and the single wrong is worth more than a pass: mop-heart's 0.3-beat pad puts time zero off-lattice (the 0.07-second lead correction makes this worse, and UNCORRECTED time zero was right there) and thins the chord histogram under the ratio at the same time, leaving a lone one-point section vote holding a wrong base. Both failures are detectable in the app at decision time, so the rule ships as: when time zero is off-lattice AND the chord ratio is under threshold, keep the declared grid or ask.

The End Witness Is Declined, and the Stand-In Resolves

The app's note on Study 31d had diagnosed the end witness's universal ritardando abstention as an eighth-note artifact: a half-beat gap rounds to one beat and halves the local tempo estimate. The bench implemented the app's fix as a second version beside the first, with tests, and measured it. On the real songs it still abstains on all twelve reads, now as a measurement rather than an artifact, letter's ending runs eight percent short of the lattice beat and mop's twenty-two percent, swing intervals passing the filter and dragging the median. On steady synthetics it votes faithfully, but it votes the LAST ONSET'S class, and these songs end on a beat-four snare, so the assumption that songs end on a downbeat fails even in the cleanest case; its one wrong vote on the waltz's correct base was the sole humility-gate violation of run two, teeth doing their job on the applicant. The weight-one seat the app hoped for is declined on the evidence. The re-draft arm closed the loop: drafts on the one-beat pads, bars placed by the voted class under the 31c rule, scored 72 of 76 on letter, exactly its chart-grid reference, and 152 of 153 on peaceful, above its reference. The registered stand-in for the missing fifth and sixth songs resolves as: humility gate passed by every shipping witness, no false moves on synthetics, one named real-world case, and the real songs remain the confirmation when they arrive. Family spend 16 cents of a 30-cent fence.

Dated Note: the Endings Do Not Slow Down

Published the same day after Kevin corrected the ritardando reading: no corpus song slows at the end. The one Badverbs song that genuinely does, Interbellum, is not in the corpus. The bench measured the endings and the band is right: eat's last bars are metronome-exact eighths at 0.49 to 0.51 beat, letter and mop end in dense sub-beat fills, peaceful plays eighths, rests five beats, then lands its final hit 0.03 beat from the lattice, and every song rings 0.2 to 0.4 seconds past its last strong onset. A true ritardando would show gaps longer than a beat and growing; none exist, consistent with the metronome-exactness correction of November 7. So the end witness's abstain label is corrected from ritardando to busy ending, and its decline stands for the sharper reason: the local-beat estimator, in both versions, cannot distinguish a busy ending from a tempo change, and these recordings hold their click to the last bar.