The Appleseed Cast tab work, audited

Every claim checked against the files and the live services, 10 August 2026. Note counts read by parsing each .gp5. Stem levels measured with ffmpeg astats. Songsterr state read from its public API and song pages.

Is it fraud?

No. One real fabrication, caught and repaired the same afternoon.

The published page matches the files exactly, 0 mismatches across 40 numeric rows, and it prints silent rather than a fake number on all six empty lanes. A rebuild started at 1:21 PM did write bass over three near-silent stems. It was found at 1:26 and repaired by 1:49, and both the gate and the generator are fixed.

Do all the songs have guitar, bass and drums?

No. 46 of 56 songs, and 6 of those carry an empty lane.

Two Conversations (2003), 10 songs, has no all-parts file at all. Five songs ship with no bass, one with no drums.

Are some missing vocal?

All of them. 72 files, 0 vocal tracks, ever.

40 of 46 songs have a real vocals stem on disk, median -27.7 dBFS and 33 percent alive. The word vocals appears once in build_allparts.py, inside the LANES tuple on line 100, and is never read again. A Songsterr AI tab for this band ships five parts, two guitars, bass, drums and the vocal line as a Tenor Sax track. The rebuild ships three or four, and the vocal is the one it drops.

Are they synced to YouTube?

No. Zero of the 46 files are on Songsterr.

Nine tabs sit in the account, six of them drums only. The upload step is still open in STATE.md, waiting on your authorisation.

The numbers

46all-parts files

56songs on the 5 records

10not built at all

5empty bass

1empty drums

0on songsterr

9tabs in your account

6of those, drums only

264-track rebuilds

0files failing the gate

0vocal tracks written

40songs with a live vocal stem

The fabrication, found and repaired the same afternoon

Caught at 1:26 PM, repaired by 1:49 PM. batch_twoguitar.py, PID 44900, started at 1:21 PM building a four-track version of each song whose guitar stem splits into two parts. It shipped with no level check of any kind, and the reality gate verify_allparts.py globbed *-ALLPARTS.gp5 only, so every *-4TRACK.gp5 it wrote passed unexamined.

This was the same defect the 9 August pass found and fixed for drums, where Messenger took 319 drum notes over a kit measuring below -100 dBFS. It came back for bass, because the new script never inherited the gate.

Bass written over near-silence, then removed

Song	bass notes written	bass stem RMS
The Immortal Soul of Mundo Cani	330	not measured
Convict	1,247	-72.8 dBFS
A Tree for Trials	205	-71.9 dBFS

The control, same script, same run

Every song whose bass stem carries real signal came back within about 10 percent of the three-track count. The separation between the two groups is roughly 50 dB, which is what made the three bad files identifiable rather than a matter of taste.

Song	bass notes written	stem peak	stem RMS
On Reflection	1,415	-4.1 dBFS	-20.4 dBFS
Blind Mans Arrow	394	-6.2 dBFS	-21.9 dBFS
Messenger	56	-31.2 dBFS	-52.5 dBFS
Doors Lead to Questions	150	-6.2 dBFS	-20.9 dBFS
Steps and Numbers	896	-7.4 dBFS	-21.1 dBFS
Mile Marker	525	-8.8 dBFS	-23.7 dBFS
View of a Burning City	1,274	-12.0 dBFS	-25.1 dBFS

Convict is the clearest case. Its six stems measure guitar -26.5, other -20.7, piano -21.3, vocals -30.0, drums -49.5, and bass -72.8 dBFS RMS. There is no bass in that stem to transcribe, and 1,247 notes were written from it.

What changed

verify_allparts.py now globs *-4TRACK.gp5 as well, reads the three-track sidecar when a four-track file has none of its own, and accepts 2 to 4 tracks. Before the sidecar fix it printed (no stem) on every four-track row and passed them all.
batch_twoguitar.py now measures the bass stem through lane_levels.measure before transcription and writes no bass part when the lane is dead on both peak and RMS, so the bad notes are never produced. Backup: batch_twoguitar.py.bak-before-lane-reality-gate-2026-08-10.
strip_silent_tracks.py repaired the three files already written. Each keeps a .bak-before-strip-silent copy.

Gate now: 71 files checked, 0 failing.

The bass is misrouted, not absent

All five empty-bass songs measure 0.0 percent alive on their bass stem, so the strip was right about the stem. The record is a different question. Filtering each stem to 40 to 160 Hz and measuring again separates the two readings, and the control settles which one applies.

Song	bass	other	piano	guitar	reading
On Reflection	-22.4	-43.9	-48.7	-32.4	separation worked
Convict	-82.4	-34.9	-33.1	-38.7	bass misrouted
Bird of Paradise	-100.0	-45.3	-49.5	-40.4	bass misrouted

On a song whose separation worked, the bass stem owns the low end by 21 dB over other. On Convict the ordering inverts: the 40 to 160 Hz band is 49 dB louder in piano than in bass. The low end is on the record. Demucs put it in the wrong stem. Calling that bass missing is a statement about the split, so these five need re-separating before any verdict about what the band played.

The vocal was never in scope

Every output file, all 72 of them, carries the same shapes: Guitar, Bass, Drums or Guitar 1, Guitar 2, Bass, Drums. No vocal track has ever been written. Meanwhile the six-stem split produced a vocals.wav for every song, and 40 of 46 carry real signal.

Songsterr's AI writes the vocal melody as a Tenor Sax track, generated from the YouTube link. The section below sets out where that comes from and the order that keeps it.

A correction, to my own claim

An earlier version of this page said the "six songs lost their AI parts on 6 August" story was unsupported. That was my error, and it is withdrawn. I read /api/meta/{songId}/revisions returning a single Initial revision, 1 track, createdVia GP as proof that nothing had been overwritten. It is the opposite. songsterr-upload/SKILL.md records that Songsterr keeps one revision per tab and that importRevisionInput replaces the whole tab in place, which is exactly why the old parts cannot be fetched back.

The evidence that settles it: all six of those tabs carry a YouTube videoId. s6375099, s6375161, s6375346, s6375387, s6362089 and s6375400 each have one. The skill states that the import does not attach audio and that a .gp5 dropped into a blank tab has no recording behind it. A tab created by a GP import cannot have that video. Those ids were made by the YouTube plus AI transcription flow and then overwritten by the drums-only import. Cathedral Rings went from 5 tracks to 1 that way.

Where the vocal actually comes from

Songsterr's AI writes every instrument, the vocal line included, from the YouTube link. That is the whole source. The sequence in the skill is /new, tick the instrument, paste the YouTube link, click Transcribe tab with AI, wait about 8 minutes, and the result is a synced multi-track tab whose fifth part is the vocal melody.

So the vocal is not gated by what happens to be in the catalogue today. Every one of the 46 songs can have one, and the page carries a checked YouTube link for each one. Three of them need no new transcription, since a vocal-bearing tab is up:

Song	Album position	Songsterr id	What it holds
Fishing the Sky	Mare Vitalis 02	s1580	5 tracks, human
Forever Longing the Golden Sunsets	Mare Vitalis 03	s2038	6 tracks, human
On Reflection	Low Level Owl I 02	s6206980	5 tracks, Songsterr AI

The other 43 each need one AI transcription run, so the credit balance is the real constraint here. /api/contributions/available-transcriptions-count read 46 of 50 remaining on 6 August 2026 and answers HTTP 401 without a login, so the current balance needs a signed-in read.

The order that keeps every part

#	Step	Why it is in this position
1	YouTube link, then AI transcribe	The import attaches no audio. A gp5 into a blank tab is a silent tab.
2	Download the AI `.gp5`	It carries the vocal, which nothing local can produce.
3	Swap in the stem-built guitars, bass and drums	Keep every other track byte for byte.
4	Import the merged file	`importRevisionInput` replaces the whole tab, so it must carry all parts.
5	Verify after a full reload	Must read `Synced video: Full mix`, and `tracksCount` must equal what the AI produced, never 1.

Publishing stays a separate act. Every tab this makes is private and unpublished.

Every song, lane by lane

A lane reading silent holds a real track with no notes in it. The 4-track bass column is today's rebuild, where it exists.

#	Song	Bars	Guitar	Bass	Drums	4-track bass	Bass stem RMS
Mare Vitalis (2000) 10 songs
01	The Immortal Soul of Mundo Cani	93	586	silent	463	0	not measured
02	Fishing the Sky	187	1,981	709	1,700	709	not measured
03	Forever Longing the Golden Sunsets	181	3,020	1,028	1,945	1,028	not measured
04	Mare Mortis	119	2,033	740	1,759	750	not measured
05	Santa Maria	125	2,090	790	1,489	800	not measured
06	Secret	100	2,557	802	1,423	804	not measured
07	...And Nothing Less	107	2,389	869	1,290	913	not measured
08	Poseidon	143	1,888	900	1,066	not rebuilt	not measured
09	Kilgore Trout	131	1,432	519	1,202	563	not measured
10	Storms	181	3,491	510	1,940	not rebuilt	not measured
Low Level Owl Volume I (2001) 14 songs
01	The Waking of Pertelotte	74	600	silent	456	not rebuilt	-74.1 dBFS
02	On Reflection	215	4,328	1,362	2,464	1,415	-20.4 dBFS
03	Blind Mans Arrow	83	1,918	383	771	394	-21.9 dBFS
04	Flowers Falling From Dying Hands	175	739	95	29	not rebuilt	not measured
05	Messenger	16	90	16	silent	56	-52.5 dBFS
06	Doors Lead to Questions	58	584	139	1,032	150	-20.9 dBFS
07	Steps and Numbers	192	3,068	896	1,912	896	-21.1 dBFS
08	Sentence	148	1,861	383	37	not rebuilt	not measured
09	Bird of Paradise	120	1,127	silent	171	not rebuilt	-63.5 dBFS
10	Mile Marker	92	2,252	496	1,536	525	-23.7 dBFS
11	Convict	278	1,007	silent	586	0	-72.8 dBFS
12	A Tree for Trials	36	308	silent	348	0	-71.9 dBFS
13	Signal	83	1,722	577	1,017	not rebuilt	not measured
14	View of a Burning City	256	3,109	1,274	1,870	1,274	-25.1 dBFS
Low Level Owl Volume II (2001) 12 songs
01	View of a Burning City reprise	71	487	217	610	231	not measured
02	Strings	254	2,657	642	1,319	749	not measured
03	A Place in Line	128	2,797	719	1,063	759	not measured
04	Shaking Hands	70	345	230	165	370	not measured
05	Rooms and Gardens	228	4,535	1,108	1,971	1,160	not measured
06	Ring Out the Warning Bell	195	1,310	861	2,191	879	not measured
07	Sunset Drama King	192	3,453	663	1,752	754	not measured
08	The Last in a Line	152	2,285	150	597	not rebuilt	not measured
09	Decline	28	483	120	418	not rebuilt	not measured
10	The Argument	132	2,706	549	1,326	553	not measured
11	Reaction	48	1,046	365	1,091	370	not measured
12	Confession	408	4,036	1,102	2,313	not rebuilt	not measured
Illumination Ritual (2013) 10 songs
01	Adriatic to Black Sea	243	2,676	837	2,308	not rebuilt	not measured
02	Great Lake Derelict	159	2,032	874	2,500	not rebuilt	not measured
03	Simple Forms	97	1,303	225	835	not rebuilt	not measured
04	Cathedral Rings	179	2,399	747	2,382	not rebuilt	not measured
05	30 Degrees 3 Am	172	2,186	779	1,887	not rebuilt	not measured
06	Branches on the Arrow Peak Revelation	105	1,736	271	1,313	not rebuilt	not measured
07	Barrier Islands (Do We Remain)	157	2,449	808	2,169	not rebuilt	not measured
08	North Star Ordination	234	2,668	772	2,453	not rebuilt	not measured
09	Clearing Life	202	2,915	518	2,083	not rebuilt	not measured
10	Illumination Ritual	171	1,544	311	1,679	not rebuilt	not measured

What is on Songsterr

Nine tabs, all private to the account. None of them is one of the 46 all-parts builds. A YouTube video is attached to the four checked below, which is what makes a tab play against the record. The two four-track tabs return HTTP 403 without a login, so their sync state is unread rather than absent.

Song	Songsterr id	What it holds	YouTube attached
The Immortal Soul of Mundo Cani	s6285808	4 tracks, all instruments	403 without login
Fishing the Sky	s6375099	1 track, drums only	not read
Forever Longing the Golden Sunsets	s6375161	1 track, drums only	not read
Mare Mortis	s6375346	1 track, drums only	yes, nAdR6bAecQA
Santa Maria	s6375387	1 track, drums only	yes, YiIqChFcCA0
On Reflection	s6206980	5 tracks, Songsterr AI	not read
Simple Forms	s6378572	4 tracks, all instruments	403 without login
Cathedral Rings	s6362089	1 track, drums only	yes, JEfi-uzS6E0
30 Degrees 3 Am	s6375400	1 track, Songsterr AI	yes, IY6-3ltSk1M

What is missing

Two Conversations (2003), 10 songs, no all-parts file

The live page scopes five records and counts 56 original tracks. 46 were built. These 10 were not.

Hello Dearest Love, Hanging Marionette, Ice Heavy Branches, Losing Touching Searching, Fight Song, Sinking, The Page, Innocent Vigilant Ordinary, How Life Can Turn, A Dream for Us

Four more studio records never touched

The band has nine studio albums. Four were built here. These five were not, and Two Conversations is the one the page counts but does not deliver.

The End of the Ring Wars (1998)
Two Conversations (2003)
Peregrine (2006)
Sagarmatha (2009)
The Fleeting Light of Impermanence (2019)

One stale claim in the notes

STATE.md publishes a table showing Bird of Paradise bass: 1 before, 522 after, presented as the outcome of the octave-fold fix. The shipped file holds 0. The fold did restore 522 notes, and strip_silent_tracks.py emptied the track 27 minutes later because the stem measures -63.5 dBFS. The live page prints silent for that cell and is correct. The note was never updated.

What is left

Done today

Gate extended. verify_allparts.py covers *-4TRACK.gp5, falls back to the three-track sidecar, and accepts 2 to 4 tracks. 71 files checked, 0 failing.
Generator gated. batch_twoguitar.py measures the bass stem through lane_levels.measure and writes no bass part for a dead lane.
Three files repaired. Convict, A Tree for Trials and The Immortal Soul of Mundo Cani, each with a .bak-before-strip-silent copy.

Still open

Re-separate the five dead bass lanes before calling the bass missing for good. Convict's low end may have gone into other at -20.7 dBFS rather than into bass, which is a separation failure and not an absent part.
Correct the Bird of Paradise row in STATE.md. It still publishes 522 where the file holds 0.
Build Two Conversations, or say on the main page that the delivered scope is four records rather than five.
Authorise the Songsterr run. Nothing reaches YouTube until then. Order matters: the YouTube link first, then the .gp5 import.

Sources. Note counts: PyGuitarPro parse of each file under ~/Projects/_outputs/appleseed-allparts/allparts/. Stem levels: ffmpeg astats on the demucs six-stem output. Songsterr: /api/songs?pattern=appleseed cast and each song page, read 10 August 2026. Live page: drwu-htmls.vercel.app/the-appleseed-cast-songsterr, HTTP 200, 82,611 bytes.