Study 15 · Listening
The patch registered at a larger n, the ratio's margin collapses to two hundredths, and the local-cost clause restores it
Listening, the registration
Study Overview
The Second Gated Song, and the Registration
The two remaining asks from Stem Shovel's Study 14 response, run the day the band revised the Peaceful Dreams chart to machine-true form: all ten sections grid-match its declared starts exactly, the first fully machine-true chart in the catalog, which qualifies it as the second gated song. The port is verbatim at the app's bc683871 (v0.201.2: the bench-recommended 1.6 constant, audibleSpan, and the FIRST-ONSET and LAST-SOUND edge rules), with the parsers reused from the Study 14 pins. The corpus: three real songs with two gated, six real wrong cross-pairs (every recording against every other song's chart), and the six synthetic probe pairs for context. Zero spend.
The Constant Survives by Two Hundredths
G1 passed, barely. The three right-chart ratios land at 2.04, 3.21 and 2.53, and all six wrong cross-pairs stay below 1.6, but the two new pairings involving Peaceful Dreams land at 1.57 and 1.58, a margin of two hundredths. The constant survives its first expansion of n and is officially thin: same-key, same-tonality wrong pairings close most of the gap the Study 14 pairs showed.
The Deliverable: the Second Clause, Stated by Number
G3 passed clean, and it is the study's deliverable: the second same-piece clause the app asked for, stated by number exactly as the 1.6 was derived. The real-domain median per-bar local costs separate completely, right charts at 0.172 to 0.243 against wrong charts at 0.365 to 0.482, so the recommended clause is median local cost at or below 0.30, AND-ed with the ratio. The complementarity is the headline: the ratio's two 0.02-margin near-misses carry medians of 0.365 and 0.400, comfortably caught, so the two clauses fail in different directions and together restore the margin the ratio lost. Disclosed per the home-field rule: three synthetic wrong pairs would pass both clauses, in a domain production never sees.
The Gate That Failed by Its Own Letter
G2 failed, and the anatomy splits three ways. First, the edge rules work where they were aimed: letter-lands' intro miss of 3.61 seconds goes to 0.00 and its outro miss of 9.11 to 0.08, and peaceful's intro lands at 0.00. Second, peaceful's outro lands 3.33 seconds late against the 1.0-second gate: a 20-bar single-chord decay rings out above audibleSpan's 2 percent floor, inflating the LAST-SOUND anchor, a real and bounded residual left open with the app as one more refinement. Third, the bench mis-drafted its own interior clause, writing a no-regression intent as an absolute ceiling that letter-lands' pre-existing stragglers never met in any prior study; the data shows zero interior regressions against Study 14's values, and the gate fails by the letter the bench wrote. Registration lesson seven: a no-regression clause must name its baseline, not restate a hope.
Verdict, Scorecard, and the Queue
The scorecard: 1 right, including eat-all-the-clocks near 2.5; 2 half, with both intros and letter's outro repaired exactly as traced while peaceful's ring-out residual and the mis-drafted clause went unpredicted; 3 right in full, including the margin claim. The verdict delivered: v0.201.2's constant and edge rules are registered as validated at n of 3 right and 6 wrong with the margin stated; the local-cost clause is no longer optional belt and braces, it is what restores the same-piece check's margin; and the LAST-SOUND refinement for long ring-outs stays open. Study 16 is queued by the app's chart-idiom response: re-running on the band's five stored chart texts through the production parser once that work lands, with one point of record corrected on the way, since their Peaceful Dreams eight-bar audit predates the band's revision, whose stored text now parses 152 bars clean.