The Appleseed Cast · built 31 July 2026, corrected 6 August 2026

Illumination Ritual: drum tabs for Songsterr, whole album

All ten tracks built from DrumSep 5-piece splits through /stems-to-guitar-pro-drums. Every file passed the same 19-point render gate.

10 / 10

tracks with a drum tab

19,609

notes written

1,570

bars

10 / 10

passed all 19 gates

0. What changed on 6 August 2026

Every number on this page was measured on 31 July, before the flam fix of 4 August and before any phantom hit was removed. Nothing regenerated the page, so it kept publishing counts for a file that no longer existed in that form. Three generations of the same twenty songs were live at once: the 31 July build behind these tables, the 4 August flam fix in the output folders, and a phantom strip that had reached only four Mare Vitalis songs. The remaining sixteen were run through the phantom pass today, all twenty are now filed in one folder, and every count below is read straight out of the shipped .gp5.

109 notated hits came out of this album, and the flam count fell hardest, because the first pass wrote grace notes onto rolls a drummer cannot play.

Song	Notes before	Notes now	Cut	Flagged	Lane coverage
01 Adriatic to Black Sea	2,309	2,308	1	1	both lanes checked
02 Great Lake Derelict	2,524	2,500	24	24	both lanes checked
03 Simple Forms	836	835	1	1	both lanes checked
04 Cathedral Rings	2,384	2,382	2	2	both lanes checked
05 30 Degrees 3 Am	1,887	1,887	0	0	both lanes checked
06 Branches on the Arrow Peak Revelation	1,332	1,313	19	19	both lanes checked
07 Barrier Islands (Do We Remain)	2,191	2,169	22	22	both lanes checked
08 North Star Ordination	2,479	2,453	26	26	both lanes checked
09 Clearing Life	2,084	2,083	1	1	both lanes checked
10 Illumination Ritual	1,692	1,679	13	20	toms lane 50%

7 flagged hits were left in place on purpose. The stripper refuses to delete from a lane whose best alignment offset does not lock, because deleting on a bad alignment removes real hits. The lane coverage column gives the share of notated hits that line up at the best offset for the lane that failed; below 70% the flagged count records a suspicion the tool could not settle, and those bars want an ear before submission.

The pass covered two of the five lanes. sd_phantom.py defaults to kick and toms, and the run took that default, so the snare, hi-hat and cymbal lanes were never tested against the audio. That is 13,160 of this album's 19,609 notes, 67%, still unexamined. Every count in the table above is real; the phantom column is a floor rather than a total, and the three untested lanes may hold more.

1. The album

#	Song	BPM	Bars	Notes	Bells	Ghosts	Flams	Tempo
01	Adriatic to Black Sea	177	241	2,308	58	175	49	MEDIUM
02	Great Lake Derelict	137	149	2,500	161	148	81	MEDIUM
03	Simple Forms	144	61	835	0	42	8	MEDIUM
04	Cathedral Rings	155	178	2,382	0	241	68	HIGH
05	30 Degrees 3 Am	161	161	1,887	0	107	40	MEDIUM
06	Branches on the Arrow Peak Revelation	138	102	1,313	0	106	51	LOW
07	Barrier Islands (Do We Remain)	148	153	2,169	0	163	144	HIGH
08	North Star Ordination	155.25	217	2,453	0	196	179	HIGH
09	Clearing Life	154	171	2,083	9	136	47	HIGH
10	Illumination Ritual	168	137	1,679	25	75	30	MEDIUM

Album totals: 253 bells, 230 open hats, 0 rim clicks, 1,389 ghosts, 697 flams, 226 crashes, 2,403 rides.

The 2,403 rides are a fallback label rather than a measurement. classify_cymbals in sd_lanes.py opens with out = ["ride"] * len(F) and then overwrites only where a positive test fires, for a crash or for a bell. No branch ever writes ride, so every ride counted here is the initial value that survived. Of this album's 2,882 cymbal-lane notes, 479 carry a positive verdict and 2,403, which is 83.4%, carry the default. That is the same lane the table above marks untested, so those notes went out with neither an audio check nor a positive classification. Measured 2026-08-29, evidence in ~/Projects/_outputs/appleseed-cymbal-default-ride-2026-08-29-macbook/.

Track 04 nearly did not happen. The album's separation script skipped it on the note "already has a published 6-stem set", and that pack carries bass, guitars, keys, metronome, other and vocals with no drums file at all. A drums stem does exist, filed under _EXCLUDED_lower_quality, 277.1 s at -18.2 dBFS. DrumSep split it into five real pieces (hi-hat 71.5% air band, cymbals 21.9%, no silent lane) and its tempo came back 155 at HIGH confidence, matching the stem-service prior exactly.

2. Two corrections to what was published yesterday

The inventory was wrong. Yesterday's page said two Appleseed Cast songs had 5-piece splits. Nine did. Seven were sitting in ~/Projects/_outputs/illumination-ritual-demucs/drumsep/ behind an _DRUMSEP_ALL_DONE marker and a validated manifest, in a directory that was never opened.

Illumination Ritual was at the wrong tempo. It shipped at 145 BPM, taken from a stem-service filename. Building a real estimator for the other eight produced a test that could be run backwards on the finished one, and snare-backbeat concentration puts it at 168. The same test returns 154 on Clearing Life at 0.780 against a 0.180 runner-up, matching that song's independent prior exactly, which is what makes it trustworthy here.

3. Tempo: three estimators are wrong before one works

Milliseconds off the grid improves monotonically as the grid gets finer, so the scan runs to the top of its range. It reported 184.5 BPM.
Scale-free phase error anchored at the first onset is still wrong, because the first onset is rarely the downbeat. It scored the correct answer worse than random.
A grid-capacity filter rejected every candidate including the known-good one, because its second condition was measuring flam spacing rather than grid spacing.
Snare-backbeat concentration works. The right grid puts the strong snares on two slots eight sixteenths apart and a wrong grid smears them. Score concentration only, since the slot index is arbitrary until the downbeat is known.

Four songs came back HIGH, five MEDIUM, one LOW. Treat MEDIUM and LOW bar numbers as provisional against the audio.

4. An ornament is a minority of its lane

The first album run produced garbage that still passed every render gate.

Song	First run	Class share of lane	In the shipped file
07 Barrier Islands	506 rim, 10 ghosts	95%	0 rim, 163 ghosts
07 Barrier Islands	282 bells, 14 rides	95%	0 bells
07 Barrier Islands	513 open hats, 135 closed	79%	closed hat part restored
02 Great Lake Derelict	458 rim, 9 ghosts	95%	0 rim, 148 ghosts
05 30 Degrees 3 Am	196 bells, 27 rides	over ceiling	0 bells
01 Adriatic to Black Sea	58 bells, 175 ghosts	under ceiling	unchanged, the control

Bells, rim clicks and open hats are accents inside a part, so a class covering most of its lane is a bad cut. MAXFRAC_ORNAMENT = 0.45 in sd_lanes.py. A runaway rim class also starves ghost detection, since ghosts rank within the hits still called snare, so a collapsed ghost count is the symptom to watch for.

5. Rim clicks: zero on all ten, and that is a measurement

Every one of the ten refused the side-stick class. On Illumination Ritual the reason was the ghost confound: body energy correlated with loudness at +0.679 and 13 of 13 low-body hits sat inside the ghost class. On Barrier Islands, Great Lake Derelict and Cathedral Rings the reason was the ornament ceiling. Either the album has no rim-click parts, or this detector cannot find them on this material. It has not been shown that a rim part exists and was missed.

6. What the gate proves, and what it does not

10 songs, 10 gate-green, 19,609 notes, 1,570 bars

  PASS  every velocity on the 16-step ladder
  PASS  one fixed string slot per drum family
  PASS  kick and side stick never share a slot
  PASS  ride is GM 51, never GM 59
  PASS  zero ghostNote parentheses
  PASS  every requested ghost carries the staccato dot
  PASS  every requested flam survived as a legal grace
  PASS  per-drum counts match the event ledger
  PASS  zero hat_pair / snare_rim / crash_ride / three_hand

session_fraud_check: PASS

That proves each file is well formed and matches its event ledger. It says nothing about whether the detection heard the right thing. The 506-rim file passed all 19. Whether Guitar Pro and Songsterr draw these correctly is the check that has to happen in the editor.