1. The inventory, and the 22 events the old census missed

Every drum event in the live tab now has a stable id, a bar, a position in sixteenths, a subdivision, a score time, a lane, a staff line, ghost and accent flags, a voice, a duration and a count of simultaneous events. After three parser corrections all 248 bars balance exactly against their own time signature.

2. Both directions, every lane

Direction one asks whether each notated lane has audio that can hold it. Direction two asks whether each stem carries reliable events the score omits. The audio count is a detector setting rather than a fact, so it is swept instead of quoted once.

For every lane there is a setting where the two counts nearly meet. A threshold can be tuned to any target, so that alone settles nothing. The next section is the part that cannot be tuned.

3. Is the surplus a real stroke or residue

Take the loudest N detections, N being the notated count, and ask two questions about the rest.

Test A, the histogram

A stem carrying two real dynamic populations is bimodal. Residue is one mode with a quiet tail.

Test B, the grid

Do the events land on a sixteenth at the measured 136.7 BPM? The phase-uniform baseline is 0.4557 at 25 ms tolerance. A jitter null lands at 0.433 to 0.491, so two independent baselines agree.

4. What the notation says about itself

The file stores loudness twice, as a ghost bracket and as a sticky beat dynamic. Resolving the dynamics in notated order needs no audio at all.

Mean dynamic rank 2.227 against 4.031, a gap of 1.804 ladder steps. Mann-Whitney z = -30.58, effect A = 0.0907.

Within-bar spacings: the flagged set sits 1 or 2 sixteenths apart on 62.89% of its gaps, the unflagged set on 35.88%. Permuting the ghost label among snare events inside each bar, 2,000 draws, gives a null mean of 0.5391. Observed is 1.167x the null with 0 exceedances of 2,000 draws, which is p = 0.00050 under the same finite-sample estimator. The ratio is modest, so the right reading is strong evidence under this within-bar permutation model rather than certainty.

5. The anchor, four procedures

Whole-song kick-sequence DTW, criteria fixed before the run

The stability criterion was invalid as written. Fixed-endpoint DTW forces the remaining score across the same audio when an interior block is removed, which is why it reported 60 to 87 second disagreements. It is reported as a failure rather than quietly redesigned.

Split-half independent fits

Two maps fitted on overlapping halves disagree by 11.6 s across their 60 s overlap, with a range of only 441 ms. One fit has slipped by a near-constant lag. At 2.7 kick onsets per second that lag is about 31 events, close to the 32-event surplus.

18a2dd, the slope-gated chroma DTW

Bar numbers here are read from the Songsterr part JSON for revision 8852151, saved as raw/8.json and inventoried in score_event_inventory.csv.

The earlier chroma DTW failed, and every disagreeing run also had a broken warp slope. That filter was chosen after seeing the results, so it was selection. This run declared it first, then ran three band radii against three windows, nine cells of nine runs each.

04ae20, the three parts that were skipped

The lead part clears, which gives the second passing external source the earlier run said it lacked. The two then disagree: lead puts bar 102 at 198.197 s and bass at 200.870 s. The harmonies row is unevaluable, since a chance rate of 0.960 leaves no room for a result.

02a84f, the calibrated null

The 1.5x rule compares an observed rate to an analytic baseline and says nothing about what the fitting procedure manufactures on its own. Each draw destroys the correspondence, then reruns the whole procedure including the train and test split.

Every p is the finite-sample estimator (r + 1) / (n + 1) over 600 draws. A permutation p of exactly zero is not a possible value, so bass with 0 exceedances of 600 reads 0.0017 rather than 0.

6. Why a passing gate still places no note

All four of 18a2dd's criteria measure whether the retained runs agree with each other. None of them measures whether the runs are right.

Extending the same procedure across the song exposes it. Evaluating the identical slope gate at each window's centre instead of at bar 102 drops two windows to 0 of 9 retained and leaves one window keeping seven runs that span 2,066 ms. The gate is location-dependent.

Holding it to its own survivorship floor across a 110 s window leaves 14.0 s of usable map, which is 2.8% of the song, in two stretches: 193.50 to 197.00 s and 254.25 to 264.75 s. Six bars fall inside. The five contiguous ones were then scored under a window slid by minus one, zero and plus one bar, every shift measured on the same three interior bars so that no shift wins by dropping one.

7. The four findings

8. Bugs found in my own work

• Dot and tuplet double-count. Songsterr's duration array already carries both, so applying them again pushed 88 of 248 bars past their own signature. After the fix all 248 balance.
• A null that could not fail. The first grid test rotated onset times circularly, which preserves every inter-onset interval and returns the data. It scored 1.001x on every set including the loud ones.
• A leave-one-out that could not work. Fixed-endpoint DTW cannot hold out an interior block.
• A shift test biased by its own edges. Scoring shift +1 over one fewer bar lowered its total mechanically. Rescored on the interior bars only.

9. Completion

10. The gate the next attempt has to pass

18a2dd passed four criteria fixed in advance and was still a bar out, because all four asked whether the retained runs agreed with each other. The next pre-registration needs a criterion that measures absolute identity instead, and it is written and hashed already: PREREGISTRATION_absolute_bar_identity.md, sha256 da96a659…05d71847.

Frozen landmarks, before any fitting

At least twelve landmarks from two source families, both excluded completely from the fitting, at least six per family, covering before, inside and after the passage. Every row records score bar, subdivision, audio timestamp, annotation uncertainty, source family, and the evidence that its phrase is unique within two neighbouring bars. Fewer than that returns UNEVALUABLE, never a negative.

The integer-shift contest

Score the fitted map against shifts of minus two through plus two bars on identical landmarks and identical assignment rules. Zero shift passes only when it uniquely minimises error inside each family, its median absolute residual is under 75 ms, its ninetieth percentile is under 150 ms, the nearest competing shift is at least 250 ms worse, and at least eleven of twelve landmarks prefer it.

The target-excluded drum check

Kick, hat and cymbal only, since snare carries the disputed flags and the tom stem is contaminated. Twelve consecutive interior bars, identical events under every shift, and zero shift must win separately for every retained lane and for their pooled cost, with a block bootstrap interval favouring it. Any source family choosing another bar returns FAIL.

A pass would validate only the span between the first and last frozen landmark. Every individual flag needs anchors bracketing its own score location, so a whole-song result needs the gate run again over every flagged region.