Study 14 · Listening
The shipped alignment and reconciler validated, and the same-piece constant caught before release
Listening, the validation sweep
Study Overview
Four Asks, Five Arms, Ten Files Ported Verbatim
Stem Shovel shipped the Study 12 alignment and the Study 13 reconciler within hours of each report and sent back four validation asks. Study 14 runs all of them on verbatim ports pinned at the app's own commits: the alignment set, the reconciler utilities, and the band-markdown parsers. Five arms, about $0.40 of a $2 fence (two model calls in the reconciler arm), one registered amendment.
What Validates Exactly
G1 passed exactly: the app's parsers round-trip the bench's material bar for bar and section for section (76 of 76 and 86 of 86 on the two transcriptions, 24 of 24 on a marked measure summary, and the 6/8-in-eighths convention rendering the app's bar seconds to the millisecond), which means future bench studies can feed production's own parsers instead of maintaining ports. G3's aligner clause held too: fed through its own parser path, the shipped aligner reproduces Study 12's precision, six of eight non-edge boundaries within 0.11 seconds of the band's declared starts, aligning a 141-second song in 1.1 seconds.
The Constant Caught Before Release
G2 is the study's reason to exist. The shipped same-piece constant, a tritone-probe ratio threshold of 3, was set from synthetic numbers where right charts score 24 to 49. On the real recordings, right charts score 2.04 and 2.53, both below 3, so samePiece returns false on the band's own correct charts, every placement gets hedged, and the song-check card would tell the band their chart does not seem to match their own recording. Amendment 1 measured what the registered pairs could not see, the real-domain wrong charts: 1.17 and 1.24. On real audio the control separates cleanly, wrongs at or below 1.24 and rights at or above 2.04, just not at 3. The recommendation went back at about 1.6, the real-domain midpoint, with the caveats owned in the report: two pairs per class, one band, and a 1.6-times margin where Study 12 had an order of magnitude. The home-field lesson's third appearance, and this time caught before a release instead of after one. Study 15 has since run the shipped constant at three right charts and six wrong pairs: it survives, by two hundredths, and the local-cost clause flagged here as the second check is what restores the margin.
The Reconciler Needs Nothing, and the Edges Are Measured
G4 passed with a bonus: under production's shipped thinking configuration, the reconciler scores 44 of 76 with 15 fixes and zero breaks, beating the bench's thinking-off measurement of 41 in half the output tokens, because adaptive thinking absorbs the per-bar narration that forced Study 13's cap amendment. Zero breaks now stand across every reconciler run in two studies, and the verdict on the reconciler is: leave it as shipped. Arm D measured the two edge rules the app asked about: the first-onset snap repairs both real intros outright (7.15 and 2.41 seconds to 0.00), and the last-sound rule repairs the 9.11-second outro miss to 0.03 seconds while inheriting chart quality exactly as feared (it would worsen the song whose outro length the bench's transcription guessed wrong), so both ship gated on samePiece and hedged.
The Loop, Four Times Around in 36 Hours
The scorecard: 1 right; 2 wrong in the direction that matters, since the real-domain collapse of the ratio was unpredicted and the amendment exists precisely because the registered pairs could not see it; 3 half; 4 right; 5 half. And the thread note this page exists to carry: request, ship, validate, catch, fix shipped, four times around the loop in under 36 hours. Stem Shovel has already taken the recommendation, the constant now sits at 1.6 in production, and their engineering log carries the standing rule the catch earned: a constant gating real recordings is provisional until measured on real recordings.