Knowledge sweep · 30 July 2026

Drum tabs: what we learned, and where it lives now

Every lesson from the Sean song drum-tab work, gathered from four knowledge layers and folded back into the skill that builds the file. This page is the coverage map.

The finding. The knowledge was split across two skills, and the wrong half was in the builder. /drum-dynamics-from-audio was built 2026-07-27 and holds most of it. /stems-to-guitar-pro-drums was built one day earlier and never received the 07-27 discoveries, so it still contained the exact bugs the later session paid to find. Running it today would have produced a Guitar Pro file that opens as rests, with rides drawn as high-F notes and rim clicks on the kick line.

historical bugs the builder still carried

properties the new render gate asserts

memory files swept

sessions read end to end

1. Where the knowledge lives now

Location	Holds
`_shared/drum_gp_spec.py`	New. The encoding spec both drum skills read: ladder, staff slots, feet, grace, ghost. Run it directly and it asserts the two copies still agree.
`stems-to-guitar-pro-drums/`	Builds a new gp5 from stems. `sd_detect.py`, `sd_writegp.py`, `sd_verify.py`.
`drum-dynamics-from-audio/`	Corrects the touch of a score whose notes are already right. Never moves a note.
`songsterr-tab-guide/`	Builds the per-band guide page. `st_drumtab.py` renders a drum-tab analysis page.
memory dir	`reference_songsterr_gp5_drum_map` plus 16 related feedback files.

2. Coverage: every lesson against the builder

The middle column is what the skill said before this pass.

Lesson	In the skill?	Now enforced by
Velocity snaps to the 16-step ladder (15/31/47/63/79/95/111/127)	No, it said 20-127	writer snaps, gate fails off-ladder
Ride is GM 51. GM 59 draws as a round high-F note	No	writer remaps 59, gate fails on any 59
Ghosts get the staccato dot, never the parentheses	No, it wrote parentheses	writer + gate
Flams are `GraceEffect(duration=64)`	No, it passed 1	writer + gate on legal durations
One fixed staff slot per drum family, 7 strings	No, slots wandered	`SLOT` map + gate
Kick and side stick never share a slot	No	gate
Two arms. Feet are kick and hat pedal only	No	`limb_conflicts()`
Quantize last, or the grid deletes the ornament	No	SKILL.md build order
Run the controls before claiming a flam or a hat	No	SKILL.md, with the measured false-positive rates
Detect replacement events, not only added transients	No	SKILL.md
One writer, regression bars, counts carry definitions	No	SKILL.md
Songsterr strips articulation and re-quantizes on import	No	SKILL.md
A round-trip is not render validation	No, it cited one as proof	`sd_verify.py`
DrumSep 5-piece beats a single full-kit stem	Yes	already present
Never resample long stems, librosa is broken here	Yes	already present

3. What each rule cost when it was broken

Rule	The failure it caused
Ladder velocity	A velocity of 431 passed a full PyGuitarPro round-trip and opened in Songsterr as nothing but rests. A whole delivery cycle and a done-claim went out on it.
Ride GM 51	Ninety minutes guessing, while Songsterr's own Drum Key tab held the answer. The fix was then graded as a regression.
Ghost dot	One file carried both notations at once, 72 parentheses and 22 dots, so an audit scored it as holding zero ghost notes when it held 94.
Grace duration 64	The bar-5 flam was added and lost across four versions. It is 59.1 ms, exactly a 32nd at 127 BPM, and the container held only 16ths.
Kick and rim apart	Rim clicks drew on the kick line, and Brandon had to say it three separate times.
Two arms	Three-hand bars shipped, graded green by an inherited rule that counted hi-hats as feet.

4. Brandon's own words, and the rule each one set

He corrected a handful of bars by hand and said plainly they were examples, so each is a rule to apply across the file. Quotes are verbatim, typos included.

i didn't fix everything, just a few examples to teach you to listen better building the skill and applying it the rest of the way

What he said	The rule it sets
you put it as 3 + hits but I only have two arms	Two hands, two feet. Feet are kick and hat pedal only, so a closed hat is a hand.
drums at 76 and other places those are rim clicks · do you have them as kicks or what · youre still doing it on the kick instead of the rim	Side stick is GM 37 on its own slot. Rim clicks run through to bar 84.
some of those are the note f, the higher f, but it should be the 'x' for a ride edge or thin crash	GM 59 draws at the hi-mid-tom slot as a round note. Ride is GM 51.
when you do the note with parentheses is hard to not hit it, just add the correct dot to it · parenntheses used to be let ring not ghost	Ghosts get the staccato dot plus velocity 31.
lots of bells like bar 8 missed	A bell often replaces a cymbal at the same position. Search for class changes.
make the most correct version lose nothing	Diff each version against the last, not only against the audio.
me doing it for you defeats the purpose · go to the audio figure it out	Do not hand the work back. The data was on disk both times.

5. The five measurement traps

The offset is not what an STFT tells you. A 2048-sample window is 46 ms wide, so band energy rises when the transient enters it. That reads about half a window early: 60 ms reported where the truth was 76.
Attack strength is a rise, never peak-minus-floor. In dense playing the window before a hit sits inside the previous hit's decay. Peak-minus-floor put 22% of ride hits at zero. The rise metric put 0% there.
Fix bleed by choosing the band. A correction factor removes real accents along with the bleed.
Ghost decisions are local, velocity is global. Velocity ranks across the whole lane. Ghost and accent rank within four bars, because the ear hears an inner tap against its neighbours.
A bell stands out when the surrounding rides come down. Raising instead pinned 7 of 12 bells to the 127 ceiling, which is the same flatness in a new place.

6. Proof

The negative control is the part that matters. I ran the pre-fix writer through the new gate, and all eight historical bugs reproduced and were caught, so the gate is doing work rather than passing everything.

$ python3 _shared/drum_gp_spec.py
PASS  drum-dynamics-from-audio/dyn_core.py matches the shared spec

$ sd_writegp.py ev.json final.gp5 && sd_verify.py --events ev.json --gp5 final.gp5
  ride59_remapped_to_51: 1     ghosts_as_staccato_dot: 16     flams_as_32nd_grace: 1
  19 passed, 0 failed            EXIT=0

$ sd_verify.py --events ev.json --gp5 old.gp5        # the pre-fix writer
  FAIL  every velocity on the 16-step ladder   offenders=[431]
  FAIL  ride is GM 51, never GM 59             ride59=1
  FAIL  zero ghostNote parentheses             parentheses=16
  FAIL  grace durations are legal              offenders=[1]
  FAIL  one fixed string slot per drum family  offenders={36:[1], 42:[1,2,3], 37:[3]}
  FAIL  staff has a slot per drum family       strings=6
  FAIL  every requested ghost carries the dot  asked=16 dots=0
  FAIL  per-drum counts match the ledger       ride (1, 0)
  11 passed, 8 failed            EXIT=1

anti_ai_voice_gate   PASS on all three new files
session_fraud_check  RESULT: PASS

7. The one check nobody here can run

Whether Guitar Pro and Songsterr actually draw the file is not checkable from this side. Re-parsing your own output with the same library proves the bytes parse and nothing more, because the library reads sequentially and skips exactly the validation the real app performs. Load the gp5 in the Songsterr editor and look before Submit.

8. One flag left unfixed on purpose

The anti-AI voice gate flags Brandon's own quote at bar 9 we ride the crash as a soft-metaphor verb. It stays. The gate governs my prose, and rewriting what he said would destroy the evidence the lessons file exists to hold. Every other line passes.