On the true grid the chartless draft reaches the chart: parity on letter-lands, above baseline on peaceful-dreams, and the two gate failures name the real remaining problems

The chart-draft pipeline, corrected

Study Overview

The Registered Design and the Amendment

Arm B4 drafted each song once (same prompt, model and parameters as Studies 23, 23b and 24) on a grid built from the declared period and an anchor derived in-run: a uniform grid from the alignment start, the shipped reader's hints over it, and the whole-bar shift whose index pairing maximizes the hint root score against the chart. The brief discloses this as an oracle construction, the upper bound for the chartless feature, since it resolves whole-bar ambiguity using the chart itself. Gate one asked for draft roots within 5 points of the chart-given baseline on the three arrangement-matched songs; gate two asked the same of eat-all-the-clocks over its playable prefix, registered as a genuine open question. One defect surfaced and was amended the same day, before reporting: on eat the shift search chose a non-physical anchor, bar one 7.6 seconds before the audio, on a one-root hint margin, and the bar-zero clamp then produced a non-monotonic grid head. The void first run is preserved and disclosed, the fix (candidate anchors more than one bar before the audio are excluded) was registered with a unit test, and only eat re-ran. The near-tie itself, 56 against 55 hint roots for 'the recording skips the chart's first four bars' versus 'it opens on bar one', was read as a corollary of that song's arrangement mismatch; the second same-day correction relocated that mismatch to the bench's chart fixture.

Parity, and What Each Failure Means

letter-lands drafted 72 of 76 roots, exactly its chart-given baseline, on a grid anchored at t near zero where two studies of ruler-anchored arms had scored 36 and 52. peaceful-dreams drafted 152 of 153, one above its baseline, with 12 of its 62 dialect bars written exactly. mop-heart missed its gate at 103 of 140 against 131, and the anatomy is pure anchor phase: the oracle anchor inherits the alignment start's sub-bar residue, 115 milliseconds off the onset-comb phase, which at 224 bpm is about 43 percent of a beat, so the bins straddle the notes. Study 24's comb-phase arm on the same song, same period, drafted 136 of 140, above baseline: the whole-bar shift search cannot repair a sub-bar phase error, and the hint-maximizing anchor is not the draft-maximizing anchor. eat-all-the-clocks failed its prefix gate at 49 of 71 against 55 of 71, which was first read as answering the registered open question. The second same-day correction reopened and then closed it differently: the live chart (chartVersion 131) has 71 bars, each chorus 8 where the bench's fixture carried 12 and the outro 1 where it carried 4, so the mismatch was the bench's. Study 25b drafted eat on the corrected chart: 56 of 69 roots against the hints' 65 of 69, a registered gate FAIL, and the same 81 percent the old grid-given baseline scored, so eat's drafting ceiling is song difficulty, not grid. Predictions: one wrong (the mop miss), three right (the dialect floor, the music-start estimator's bar-one placements of 0.12, 0.12 and 0.39 seconds with mop missed by 3.7 bars, and spend), and the open question resolved to its disclosed FAIL reading.

The Measured Recipe

Across Studies 24 and 25 the corrected grid story is now measured end to end, and the recipe the bench hands the app is short: use the declared bpm as the period, it is exact; take the bar phase from the onset comb; place bar one near t equals zero. The better arm per song reaches 72 of 76, 152 of 153 and 136 of 140 against baselines of 72, 151 and 131, parity or above on every arrangement-matched song. The open problem that survives the second correction is beat-level anchor precision at fast tempos; the apparent arrangement mismatch was the bench's own chart fixture, and the standing rule it leaves is to verify the fixture against the live chart before blaming the band's arrangement.