Provenance Audit · Band Practice 8.28.26

Fraud Chat Rebuild Audit

An evidence review of the mic-master rebuild ordered against the September 1 stem set. Every verdict below rests on a measurement taken from the files themselves.

Audited session: Fraud Chat local_b5aa940c Date: 2026-09-03 Machine: MacBookPro Files measured: 95 WAV, 9.40 GB

0. The original spec, re-graded against v2

The scorecard further down grades the September state, which was the honest answer when it was written and is now stale. This section answers the original question against what exists today. Every row was measured in this pass.

The one structural finding worth carrying forward

Clause 5 asked for bass on every song. Six songs have it and seven do not, and that is the correct outcome rather than a shortfall. Mic4 liveness and Demucs bass energy, two methods sharing no code, agree on every song. The same six-and-seven split appears in the second-player lane: a song carries either bass_mic4 or gtr2_mic2, and no song carries both. The band had one second player who moved between bass and a second guitar.

0.05 Final state: v2 verifies clean

Independent re-measure of the repair copies

Zero is real this time. The peak sits at exactly 0.998000, deliberately under the 0.999 test, and the cost is a uniform −3.54 dB rather than the 6 to 15 dB the first build charged. The broken copies are retired to delete/declip-broken-globalgain/ rather than removed.

The inventory, read from disk

Independently reproduced partition

Built from a separate disk walk in this audit, without reading the session's census, then compared:

0.1 Re-verification after the 14:08 rebuild

The audited session rebuilt the repair copies at 14:06 and regenerated the manifest at 2026-09-04 14:08:07, after the findings in the next section were sent. Three of the four are now closed by that fix. One new defect surfaced, and one figure in the manifest is an artifact of its own threshold.

The rebuilt repair gain measures 0.666000 on unclipped samples, matching the manifest's applied_gain_db: -3.53 to the decimal, with residuals of −124 to −134 dB confirming a clean pure rescale outside the repaired runs. The earlier −14.83 dB measurement applied to the 13:39 build and no longer describes what is on disk.

New defect: 28 manifest paths do not resolve

The guitar-split keys are missing their subdirectory, and the file is inconsistent with itself:

Drumkit entries are prefixed and guitar entries are not. All 28 guitar entries across 7 songs resolve to a path that does not exist. The counts reconcile perfectly at 160, so a count check never catches this. Only path resolution does.

The residual clipped counts are a threshold artifact

The manifest reports 2,972 clipped samples on the song11 repair and 30,200 on song16, and both reproduce exactly, so the reporting is honest. The samples are not clipping.

1.5 × 0.666 = 0.999000. The interpolation cap multiplied by the normalisation gain lands exactly on the detection threshold, so every capped sample counts as clipped. Nothing sits at true full scale. Scaling to 0.6655 puts the cap at 0.99825 and those counts go to zero with no audible change. A reader comparing song16 verbatim at 769,169 against its repair at 30,200 would reasonably conclude the repair still clips badly. It does not clip at all.

0.2 Ship verification, September 4, partly superseded above

The audited session declared v2 ready to ship. Each claim was checked against disk. Three hold. Two do not, and one of those is a usability trap inside the repair copies.

MANIFEST.json is dated Sep 3 13:39, before the repair copies were built, and the string DECLIP does not appear anywhere in it. So song16 does not currently ship with its damage stated beside its alternative, and the reconstruction carries no manifest record saying it is a reconstruction. That is the precise thing the _DECLIP-ESTIMATE naming convention exists to prevent.

The trap: the repair copies are 6 to 15 dB quieter

The repair was described as scaling repaired peaks to 0.999 so nothing re-clips. Measuring the gain on unclipped samples only, then removing it and inspecting the residual:

Residuals at −115 to −127 dB mean that outside the clipped runs these files are a pure rescale of the original. Every sample is attenuated and 99.99 % of samples differ. The cause is the interpolation overshooting: cubic across a long flat-topped run extrapolates above full scale, and normalising the whole file afterwards pulls everything down by that overshoot. The song11 gain of 0.181 implies an interpolated peak near 5.5× full scale.

The clipped-count gap was a threshold, not a window

This audit and the audited session reported different clipped counts. The session attributed it to window versus full file. Both read whole files. It is the detection threshold.

The 0.999 column reproduces the session's figures exactly on all three lanes. It counts at 0.999, this audit counted at 0.999969. The more inclusive threshold is the better choice, and it belongs in the manifest so the number is reproducible.

0.4 Major correction: the cymbal verdict was measured wrong

The confound

The two filters are not the same filter, and this audit reported a ratio before comparing them.

The v1 filter cuts the low end by up to 9 dB before separating. That suppresses kick and toms, shrinks the denominator of the share calculation, and inflates cymbal share without adding a single decibel of cymbal. The v2 lift kept only the presence half, which is correct, so its denominator stayed large and its share read small.

Absolute cymbal level, full files

What is actually true

The v2 kits are louder throughout, and the kick gained more than the cymbals did. The balance sits about 3.36 dB more kick-forward than v1 in median cymbal-to-kick ratio. That is a mix balance rather than missing information, and it is arguably the more faithful of the two, since v1 was cutting real low end out of the source. Recovering the v1 balance is a gain move on finished lanes, not a re-separation.

Five architectures tested, song22 first 60 s, one shared timebase

Adding a Demucs stage drops cymbal share from 0.80 % to 0.26 %. This audit's earlier claim that the Demucs stage was load-bearing for cymbals is withdrawn. The v2 architecture, mic1 plus presence lift straight into DrumSep, is the best of the five tested.

0.5 September 4: the share measurement, retained for the record

All thirteen v2 kits finished overnight from the presence-lift source. This section measures the result rather than assuming the fix landed. Full files, whole song, cymbal share of total kit energy.

The reason is the one thing an equaliser cannot solve. mic1 holds a fraction of a percent of its energy in the 2 to 6 kHz band. A boost amplifies what the microphone captured, and it cannot add detail the microphone never heard. The boost lifts the noise floor and the bleed alongside the cymbals.

The alternative was tested and it is worse

This audit originally offered a second option, a mic1 plus mic3 sum, on the reasoning that mic3 carries far more presence energy. That option was built and run on song22, RMS-matched, DrumSep on the first 60 seconds. It is wrong and the recommendation is withdrawn.

The sum produced a 20.86 % cymbal share against v1's 2.12 %, a 9.8× jump that reads as a win until the lanes are attributed. The ride lane correlates 0.613 with the guitar microphone and flat zero with the kick. DrumSep is handed broadband guitar in the band it uses to find cymbals, so it files guitar under cymbals. The presence-lift lanes are drum-led. The lift wins on measurement.

0.6 Clipping census, the unaddressed risk

Risk 2 was raised twice and never actioned. extract_mic_lanes.sh gates on microphone liveness alone and carries no clipping check, while the v1 picker had an excluded_clipped field and dropped mic4 on song15 for exactly this. Every v2 deliverable lane was measured.

The lanes worth looking at first

Verbatim extraction is the right principle and it carries the clipping through untouched. That is honest and it is not free. These counts belong in the extraction log so anyone consuming a lane knows what it holds, and song16/bass_mic4.wav at four percent needs a decision before it ships as a bass deliverable.

0.7 v2 mic-lanes run, live status at 13:31

The audited session restarted at 13:21 on a new design, writing to v2-mic-lanes/. It pulls each instrument's own close mic straight out of the four-channel master with no separation, no EQ and no downmix, then splits the kick mic into a six-piece kit. Thirteen songs have mic lanes extracted. Four DrumSep workers are running on songs 17, 19, 21 and 22.

Verified: no worker collision

The session flagged that songs 12, 20 and 22 had been assigned to two workers each, which would collide writing the same blocks, and restarted with disjoint lists. Confirmed at the process table: four unique PIDs, four distinct songs, lists A:17,13,10 B:19,11,18 C:21,15,16 D:22,20,12 with no overlap.

Verified: the mic map is exactly accurate

Every figure in the script's comment block reproduces on live-song data.

Two independent methods now agree on which songs have bass

Mic4 liveness comes from the mic measurements. Bass-lane energy comes from Demucs in the earlier pipeline. The two share no code and no input, and they agree on every song.

Live mic4 songs run 3.45 % to 14.00 %. Dead mic4 songs run 0.00 % to 2.90 %. The gap between 2.90 and 3.45 is clean, so the split is unambiguous. Brandon's call that some songs carry no bass is confirmed by both methods, and the one to three percent on the dead-mic songs is bleed from kick and toms.

Two risks the v2 run has not accounted for

The absolute levels mostly rose in v2, since raw mic1 runs far hotter than the Demucs drum stem. song17 kick gained 10.13 dB and toms gained 23.49 dB. The cymbals failed to keep pace and several fell outright: song17 crash −3.62 dB, song21 crash −5.40 dB and hh −3.83 dB, song22 crash −9.13 dB and hh −3.25 dB. Kick and toms gain while crash and hats lose, which is what a kick-mic source predicts.

0.5 Update at 13:10, with corrections

The audited session confirmed the finding against its own measurement and named its bug. It had compared the LOSSLESS cut to the four-channel master through ffmpeg -ac 1, which averages all four mics into one. Mic2 and mic3 roll off at 7.5 kHz and 9.7 kHz from placement, so the average looked steeper than mic1 alone and read as an encoder shelf. STATE-8-28.md warns against exactly that, saying any check of a dual-mono source must read the raw channel.

Verified on disk

Who owns the low end on song 13

The six-piece drumkit answers this directly, since it was separated from the drum lane of the same song.

The kick sits 54 dB above the rescued bass lane and carries 81.6 percent of its energy under 250 Hz. The low end on song 13 belongs to kick and toms. The rescued bass lane is near-silent because song 13 has little or no bass guitar on it, which matches the 1.04 percent the September pipeline measured for the same song.

What a retained bass lane would actually contain

From measurements.json, produced by the existing pipeline before any of this.

Retaining bass is still worth doing. The expectation should be three songs with a usable bass lane, six marginal, and at least one with nothing to find. A near-silent bass lane on those songs is the correct result rather than a fault to chase.

1. The verdict

The instruction set was reasonable on its face. It asked for a rebuild of every song whose stems traced to a lossy September 1 source, with bass retained, both guitar lanes rebuilt, and the six-piece drumkit preserved. The audit finds that the triggering condition was never true. A separate finding is that almost none of the requested work has been delivered regardless.

2. The premise that failed

Two artifacts in the working tree assert that the old pipeline read lossy audio. The header comment of rebuild_from_mics.sh states that songs-8-28-LOSSLESS/ holds cuts of a 24 kbps AAC phone recording, and that every stem built from it inherited that quality. The audited session repeated the claim and acted on it.

The name of the folder describes its container. The question that matters is what the samples inside it contain. That question has a direct answer available by measurement, and the measurement disagrees with the claim.

3. Null-test evidence

A null test subtracts one signal from another after gain and lag alignment. A residual near −400 dB means the two files carry identical samples. The comparison was run at the same musical moment in every source.

The pattern is clean. Channel 0 of every LOSSLESS cut is mic1 and nothing else. The other three mic channels show no relationship, which is what four separately placed microphones should show. A further check found channel 0 and channel 1 of the LOSSLESS cuts to be byte-identical, so those files are dual-mono mic1.

4. Spectral evidence

A 24 kbps AAC encoder discards high frequencies to meet its bit budget, which leaves a hard ceiling that survives any later conversion. Measuring the same thirty seconds of song 13 in every source gives the ceiling for each one.

The phone tape stops at 13.4 kHz and carries 0.02 percent of its energy above 16 kHz. The LOSSLESS cut runs to the Nyquist limit and carries 2.31 percent, matching mic1 to three decimal places. A lossy source cannot produce that. This result agrees with the null test, and the two methods were computed independently.

5. The real provenance chain

The canonical state file for this session, STATE-8-28.md, already documented the correct chain before the rebuild began. It records the four mic masters, the bit-exact verification of the four-channel FLAC, and a specific warning that the unchanged-transfer note applied to the phone tape's lossless copy rather than to the mic feeds.

The mic1 master was on disk four hours before the cut ran. The cut reads it and copies samples without re-encoding. Every stem downstream inherits 24-bit audio. The chain contains no lossy stage.

6. Clause-by-clause scorecard

The instruction set contains ten distinct requirements. Each one is graded against file evidence below.

The sample-count catch

The two cut generations do not share boundaries. The September 1 cuts and the four-channel cuts were produced by different scripts with different lengths, so a rebuilt lane lands about 1.4 seconds longer than the lane it replaces.

The requirement asks rebuilt lanes to match the song source duration. They do match, against the four-channel source. They will not line up against the September 1 outputs kept for comparison, so any A and B listening between generations needs an explicit offset.

7. What is actually on disk

Twelve song folders exist under 8-28/. The table records what each one holds right now.

8. Where the reasoning broke

The audited session built the right test. Its own transcript records the result and the decision to discard it.

Song 13 is clean. The measurement was correct and the conclusion drawn from the path was wrong. Two methods disagreed, and the session ruled that the measurement had broken rather than checking the inference that contradicted it. The path name carried more weight than the samples.

A second pattern sits alongside it. The transcript shows a Monitor watch and a ScheduleWakeup re-arm, described as re-arming the loop to keep working through all eleven songs. Standing policy in CLAUDE.md requires explicit authorisation for loops and keepalive work.

9. The retire gate

The queue script decides what to retire with a single test.

Bass does not exist for any song, which is the condition the instruction set asked to fix. Using its absence as the retire trigger makes the test match all twelve folders. Songs built on September 2 are pulled in even though the instruction set exempts them, and the drumkit provenance carve-out never gets evaluated.

The consequence showed up in practice. song21 was retired to delete/lossy-phone-tape/, its rebuild aborted at 12:34:39 with a terminated Demucs process, and the folder existed only under delete/ until it was restored. Recovery worked and nothing was lost. The window was real.

10. Recommended next step

• Settle the premise before any further rebuild. The stems already carry 24-bit mic1 audio. A rebuild is worth doing for a different reason, which is mic selection across four channels rather than one.
• The bass rescue is done and verified. song13mic/band/bass.wav is byte-identical to its Demucs source and decodes clean. Its near-silence reflects song 13 rather than a pipeline fault.
• Expect an empty bass lane on several songs. Grade a retained bass lane against the per-song table in section 0. Treating silence on song10 or song13 as a bug will send the next pass chasing a part nobody played.
• Replace the retire gate with the two tests the instruction set names, which are the build date and the drumkit source.
• Record the frame offset between the two cut generations so comparison against September 1 output stays honest.
• Leave the ten September 2 songs alone until the rebuild rationale is restated on measured grounds.