jam-7-3-26-rebuild · single authority record

Ableton Sync: One Canonical Verdict

Five songs, seven mono stems each, measured against the source. Replaces the merged multi-run narrative from the last six chats.

IMPORT AT TIME ZERO

Every packaged stem shares its song's source timeline. Across song01 through song05, the summed drum pieces show zero waveform lag against the Demucs drum stem at five positions from beginning through ending, drift span zero. The bass, guitar, and drum stack shows zero lag against the source cut. Drop all seven stems at Arrangement time zero and apply no manual offset.

5
songs complete
7
mono stems / song
0
waveform lag (samples)
0
drift span

Table 1 · Canonical packaged-file synchronization

Run against the delivered mono files rather than the prepackage outputs. Five 6-second windows at 2 / 26 / 50 / 74 / 98 %, waveform cross-correlation maxlag ±500. Machine record: analysis/SYNC_FINAL.json.

songframes/chchsr7 stemssha256 ×7primary wav lagsecondary wav lagdriftonset-env
song0125,970,931144100✓present0·0·0·0·00·0·0·0·000·0·0·0·0
song029,966,600144100✓present0·0·0·0·00·0·0·0·000·0·0·0·0
song038,388,702144100✓present0·0·0·0·00·0·0·0·000·0·0·0·0
song049,040,500144100✓present0·0·0·0·00·0·0·0·000·0·0·0·0
song0510,968,993144100✓present0·0·0·0·00·0·0·0·000·0·0·0·0

Primary = summed 5 packaged drum pieces vs the Demucs drums stem. Secondary = summed packaged bass + guitar + drum pieces vs the source cut. Within each song the seven stems carry one frame count, and that count equals the source cut.

Table 2 · Reconstruction residual

Separation quality, unrelated to timing. Residual is how completely a family of stems sums back to its parent. Every lag above stays zero regardless of these figures.

songDemucs residual (cut − Σ6)DrumSep residual (drums − Σ5)note
song010.130  (−17.7 dB)0.097  (−20.3 dB)DrumSep residual 3 to 5× the others; give it a listen
song020.133  (−17.5 dB)0.026  (−31.6 dB)
song030.190  (−14.4 dB)0.018  (−34.8 dB)
song040.272  (−11.3 dB)0.030  (−30.4 dB)highest Demucs residual
song050.051  (−25.8 dB)0.030  (−30.6 dB)cleanest overall

The onset-envelope outlier, documented

Where. Prepackage test sync_check2.py, song04, 90 % window. Onset-envelope lag read −12,800 samples. The waveform lag at that same window is 0, and the 10 % and 50 % windows read 0.

Read. An onset-envelope cross-correlation needs stable transient energy. In a sparse late window it can lock onto a spurious peak. The waveform correlation there is 0, the frame counts match exactly, and the surrounding windows read 0. The correct statement is that the envelope estimate lacked stable transient information there. No shift exists. On the delivered files the canonical run reports onset-env 0 for song04 at every position.

Dataset

Source master master_4hr.m4a, Google Drive, authenticated download 2026-07-19. 49,868,049 bytes, md5 9892b41f4d84bb22, AAC 44100 Hz, 2 channels, confirmed dual mono by a full-stream L-R null test. Decoded 4:32:24, 720,768,960 frames per channel.

Pipeline: lossless 24-bit cut → Demucs htdemucs_6s → DrumSep MDX23C (5 pieces) → mono downmix 0.5*L + 0.5*R. Packages at /Volumes/T7 Shield/jam-7-3-26-rebuild/packages/song0{1..5}/.

songsource perfsource windowdurationframes/ch
song01perf020:01:03.500 to 0:10:52.410588.91 s25,970,931
song02perf361:36:14.500 to 1:40:00.500226.00 s9,966,600
song03perf944:11:20.500 to 4:14:30.720190.22 s8,388,702
song04perf050:14:14.500 to 0:17:39.500205.00 s9,040,500
song05perf080:22:02.500 to 0:26:11.230248.73 s10,968,993

Scope

COMPLETE  song01 through song05: seven mono stems each, uniform sample rate / channels / frames, five-position sync zero, SHA256 recorded, provenance written, source chain proven.

INCOMPLETE  song06 through song08: bass and guitar packaged, drum pieces not yet separated. Queued.

What this record fixed