Does re-deriving from the stem work?

Three Appleseed Cast songs, each re-transcribed from its own guitar stem with basic_pitch on the ONNX model, nothing carried over but the container. Scored on the meter clock against a cut validated leave-one-song-out. 30 August 2026.

Three songs, three different answers.

Playability improves every time. Audio accuracy does not.

Adriatic clears the cut with margin. Confession moves ten times its starting separation and stays under. Blind Man's Arrow moves the wrong way. The method is 1 for 3 on audio and 3 for 3 on playability.

3 of 3reach 0 hard failures

1 of 3clear the audio cut

7.7validated cut

EXACTround trip, all three

Every build, every number

Same tool on every row, and each song's inherited tab reproduces its published figure exactly, which is the control that says the run is sound.

build	attacks	own	chance	gain	z	separation	verdict at cut 7.7
Confession
inherited tab	2702	13.1%	13.6%	-0.4	-1.5	+0.5	under
re-derived, 16th	2705	15.4%	13.6%	+1.9	+6.8	+2.7	under
re-derived, 32nd	3059	18.0%	13.5%	+4.5	+15.3	+5.2	under
Adriatic to Black Sea
inherited tab	1764	21.3%	17.1%	+4.2	+17.7	+4.4	under
re-derived, 16th	1765	25.4%	17.1%	+8.3	+35.7	+8.6	clears
re-derived, 32nd	1945	30.3%	17.0%	+13.3	+45.4	+12.9	clears
re-derived, 32nd + 2 staves	1945	30.3%	17.0%	+13.3	+45.4	+12.9	clears
Blind Man's Arrow
inherited tab	972	24.7%	17.2%	+7.5	+14.4	+6.4	under
human Songsterr tab	581	41.0%	18.2%	+22.7	+32.8	+20.9	clears
re-derived, 32nd	1229	23.1%	16.5%	+6.6	+16.9	+5.0	under
re-derived, 32nd + 2 staves	1229	23.1%	16.5%	+6.6	+16.9	+5.0	under

Where the ceiling is

Blind Man's Arrow is the one song here with a human Songsterr tab on disk, and it is the most useful row on the page. A person reaches 41.0% own alignment on that recording. Our best re-derivation reaches 23.1%. Same tool, same stem, same clock, a gap of 17.9 points.

That gap comes from what basic_pitch detects, rather than from the measure or the grid. Adriatic reaches 30.3% and clears; Confession reaches 18.0% and does not. The songs that clear are the songs the detector happens to hear well, so the method's success is currently a property of the audio rather than of the pipeline.

Playability, uniform and finished

song	hard failures before	split	round trip
Confession	16	16 notes to a second staff	3922 of 3922, 0 pitches changed
Adriatic to Black Sea	9 at 16th	13 beats dealt into two parts	2641 of 2641, 0 pitches changed
Blind Man's Arrow	1	11 beats, 23 notes moved	1978 written, gate REVIEW both staves

The finer grid helps playability on every song tested, and the two-staff split resolves the remaining unplayable beats without moving a single pitch. That part of the pipeline is working and is independent of the audio question.

What this does not say

The audio number measures detector agreement rather than pitch accuracy. basic_pitch produced these notes from these same recordings, so a rise measures how far its note starts agree with librosa's onset detector. None of these files has been promoted into 09_MATCHED_38.

The cut is validated, and that is what makes 1 of 3 meaningful. Separation ≥ 7.7 with z ≥ 5.0, established leave-one-song-out over 162 known-wrong pairings across 54 songs: the same 7 confirm in-sample and held out, 1 false positive in 162, threshold stable between 5.1 and 7.7.

Records in ~/Projects/_outputs/appleseed-regen-2026-08-28/ under confession_rederive/, adriatic_rederive/ and blindmans_rederive/. Each song was scored by a copy of the same 54-row tool with only the case list and output path changed. The measurement bugs that invalidated the earlier −36.6 margin are documented at appleseed-audio-measure-refuted.