Three clock bugs under the Appleseed audio verdicts

29 of 46 songs were marked audio refuted on 2026-08-29 and queued for re-transcription. Three separate defects were found underneath that number, each measured against tabs whose answer is already known. Every figure here is read from the project's own tools. 29 August 2026.

1. The verdict

2. Bug one, control selection

librosa.sequence.dtw(subseq=False) forces the path to span both sequences, and the audit picks controls by onset count and ignores duration. A control shorter than the song makes the warp compress the tab into it, and the compression becomes flat runs: hundreds of attacks on one onset, all counted as hits in the 30 ms window.

3. Bug two, the reader's clock

attack_string_evidence.events() sets each bar's length to max(onset + duration) over that bar's note-bearing beats. A bar ending in a rest reports a short span, and the loss is carried for the rest of the song. Within-bar positions are fine; the bar cursor is not.

4. Bug three, the onset frame

beat_map.py takes its reference onsets from onset_detect(backtrack=True), which walks every detection back to the energy minimum before the peak. librosa.beat.beat_track puts its beats on the peaks. The two live in different frames.

5. Three clocks, same negative controls

6. The tempo maps

All 46 files were written with per-bar tempo derived from their own recordings, originals untouched, so the comparison reproduces. The grid comes from the audio alone and the notation is laid on it afterwards, because fitting a tempo curve to the tab it is about to score would repeat the circularity that let the DTW warp manufacture alignment.

7. The confirmation threshold, validated

Reading the cut off the observed maximum of the negative controls is post hoc calibration, and it guarantees zero false positives inside the same sample that set it. Brandon called that on 2026-08-29, so the cut was re-derived leave-one-song-out: for each song the threshold comes from the other 45 songs' negative controls alone, and the held-out song is judged by a number it never contributed to.

The validated standard for any promotion on this measure is separation ≥ 7.7 with z ≥ 5.0. Four machine tabs clear it: Mile Marker, Doors Lead to Questions, Steps and Numbers, and Decline. Four that the 5.0 cut had confirmed now fall below it: A Place in Line at +7.4, Sunset Drama King at +7.2, Blind Man's Arrow at +6.4, and The Last in a Line at +5.9. Three of the eight human tabs clear it, the same three as before.

8. All 46 rescored, meter clock

The configuration the evidence supports, judged at the validated cut of sep ≥ 7.7 with z ≥ 5.0. Undetermined carries no finding against the tab: 5 of the 8 known-good human transcriptions end up there too.

9. Confession, rebuilt and not promoted

Confession was re-derived end to end from 03_GUITAR_STEMS/Low Level Owl Vol II/12 - Confession.guitar.wav with basic_pitch on the ONNX model and the fretboard solved with the skill's own beam. Nothing was carried over except the container. The first build quantised to a 16th grid at 174 bpm and that grid was the whole defect: it pushed 31.0% of attacks outside the 30 ms scoring window. Requantised to a 32nd grid, 0.0% fall outside it.

Mare Vitalis track 7 is written here as And Nothing Less. The released title begins with an ellipsis and the filename on disk is unchanged.
Tools and raw records in ~/Projects/_outputs/appleseed-regen-2026-08-28/: validate_alignment_measure.py, duration_matched_test.py, banded_dtw_test.py, chance_normalized_test.py, prewarp_full46.py, reader_clock_drift.py, meter_clock.py, prewarp_meter46.py, beat_census.py, tempo_map.py, tempo_audit_full.py, write_tempo_maps.py, and confession_rederive/. Tempo-mapped files in tempo_mapped_gp/, originals untouched. Every fast path was checked against a brute-force reference before its numbers were used. Playability is unaffected by any of this: it was measured against the files rather than against audio, and the 1248 to 212 hard-failure reduction with 15 of 46 hard clean stands.