6.6 KiB
Music master verification — "RECEIPTS, M'LORD", three candidates
Date: 2026-09-24 · By: Timmy · Status: pre-verification of the claims stated in WORK-QUEUE.md Q9,
run before the independent pass was asked for and recorded so that pass has something to contradict.
Not the second opinion. These measurements were taken on the same host, with the same tools that produced the files. They can establish whether the recorded numbers are true of the delivered files; they cannot establish that the pipeline is right about anything. The independent pass is still Zeph's to make, or nobody's.
What was measured, and with what
cd /root/sovereign-music-daily/releases/2026-09-24/masters
sha256sum *.mp3
ffprobe -v error -show_entries stream=codec_name,sample_rate,channels,bit_rate,duration \
-show_entries format=duration,bit_rate,size,format_name -of default=noprint_wrappers=1 <file>
ffmpeg -nostdin -hide_banner -i <file> -af loudnorm=I=-14:TP=-1.0:print_format=json -f null - # input_i / input_tp / input_lra
ffmpeg -v error -nostdin -i <file> -f null - # full decode, rc + stderr
ffmpeg -v error -i <file> -f md5 - # decoded-audio identity
Delivered MP3s — claim vs measurement
| candidate | sha256 (of the 7,801,965-byte file) | matches STATE.json |
integrated LUFS | true peak dBFS | decode | dur (container) |
|---|---|---|---|---|---|---|
78c3791c |
75bf91ffc2a141cc564c7cbbba83a8cd82c9685001760d4444f979c6a2ee0bff |
yes | −14.12 | −0.94 | clean, rc 0 | 195.024 s |
a381d8cf |
c1cc8fda1785502378d78e23a85c7ee81faff325e780ae7f6bdf81e131aa0804 |
yes | −14.00 | −1.57 | clean, rc 0 | 195.024 s |
e6d5e7e9 |
ca24d8a2beb8715e501df5218b29dca47e2b7f93ee1747caeefe95623cc6f40d |
yes | −13.98 | −0.92 | clean, rc 0 | 195.024 s |
All three: mp3, 48000 Hz, 2 channels, 320000 bps. Decoded-audio MD5s are three different values
(b57bc3b3…, 20a02c46…, ccef58bc…), so these are three distinct renditions, not three copies of one.
Each Q9 claim, graded
| claim in Q9 | verdict | measurement |
|---|---|---|
| integrated loudness −14.0 ± 0.2 LUFS | holds | −14.12 / −14.00 / −13.98, max deviation 0.12 from the record |
| true peak ≤ −0.9 dBFS | holds as written, wording should change | −0.94 / −1.57 / −0.92. Two of three sit inside 0.02–0.08 dB of the bound, and STATE.json's own event line states the mastering target as −1.0 dBTP, which none of the three meets (−0.92 and −0.94 are above it). Quote the measured value; the −1.0 is a loudnorm parameter, not a delivered figure. |
| full decode clean | holds | ffmpeg -v error … -f null - exits 0 with empty stderr on all three |
| 195.0 s | holds, off by the encoder's padding | FLAC source exactly 195.000 s; MP3 container 195.024 s |
| FLAC-source 48 kHz stereo | holds | 3 candidate FLACs: flac, 48000, 2ch, 195.000 s |
| 320 kbps MP3 | holds | bit_rate=320000 (container rate 320041) |
| SHA-256s in the manifest | holds; and "the manifest" is STATE.json |
all three match STATE.json character for character; no separate manifest file exists. The source FLACs match logs/generated-files.json |
| "three ACE-Step CPU renders" | holds | logs/ace-result.json: engine ACE-Step 1.5, model acestep-v15-turbo, device cpu, inference_steps 16, seed 4421…, success true, started 09:50:55Z, finished 12:18:16Z, wall_seconds 8840.9 |
QC.json (the source-file measurements) also reproduces: LUFS −13.37 / −13.28 / −13.33 against the recorded
−13.3 / −13.3 / −13.3; true peaks −0.87 / −0.74 / −0.93 against −0.9 / −0.7 / −0.9. The only figure with real
spread is loudness range — I measure 11.30 / 9.00 / 12.50 against the recorded 11.1 / 9.1 / 12.1, up to 0.4 apart.
LRA is the one number here whose definition varies between implementations, so this is recorded as a method note,
not as a discrepancy: the recorded values are not contradicted.
Hook transcription — my own pass, 50–110 s window
ffmpeg -loglevel error -ss 50 -t 60 -i receipts-mlord-<tag>.mp3 -ac 1 -ar 16000 -c:a pcm_s16le /tmp/q9win-<tag>.wav
python3 -c 'from faster_whisper import WhisperModel; m=WhisperModel("small.en",device="cpu",compute_type="int8",local_files_only=True); \
s,i=m.transcribe("/tmp/q9win-<tag>.wav", language="en", beam_size=5, vad_filter=False, condition_on_previous_text=False); \
print([(round(x.start+50,2), x.text.strip()) for x in s])'
Model Systran/faster-whisper-small.en (int8, CPU), one 60 s window starting 50.0 s in each delivered master,
raw output kept at /tmp/q9-transcription.json. Wall time 96 s for the three.
78c3791c — chorus recovered at 77–91 s:
[77.00] Lord, the receipt is the proof The hats don't care who told you no
[82.00] Lord, the receipt is the proof 44 hours of tape in a row
[87.00] Say it again, the receipt is the proof
[91.00] Root around the door Keep your sure
a381d8cf — chorus at 86–105 s:
[86.00] Lord, the receipt is the proof
[91.00] The ask or receipt is the no
[96.00] Lord, the receipt is the proof
[101.00] Rude around the no, keep the show
e6d5e7e9 — chorus at 75–99 s:
[75.00] Lord, the receipt is approved
[78.00] And I don't care who told you not
[81.00] The receipt is the proof
[83.00] Forty-four hours of tape in a row
Reading it honestly. The first hook line, "the receipt is the proof", is recovered verbatim in all three
candidates, in some case three times per window. The second hook line, "the hash don't care who told you no", is
recovered in none of them: ASR returns "The hats don't care…", "And I don't care…", "The ask or receipt is the
no". The word hash never survives. Whether the render mangles it or small.en simply cannot resolve it over the
mix is not decidable by this method — that needs a listener, or a lyric-aligned wider model, not a transcript.
Verse lines around it come back close to lyrics.txt but not exact ("22,000 sets/sides on a dead post",
"Recovered from a capture nobody hopes/holds"), which is the normal error rate for sung vocals.
One artifact to not be fooled by: e6d5e7e9's window ends with a spurious "Thanks for watching!" at 108–110 s. That is a known whisper hallucination on music, not content in the file.
Limits: ASR on singing, one 60 s window per candidate, one model. This says nothing about mix quality, intelligibility to a human ear, or which candidate should be released. It is not the independent pass either — the same host made the files.