The matcher's two new rules validated on real audio: both earn their keep, and the author's own hedge is overturned in their favor

Validation, the matcher rules

Study Overview

The Handoff, and the Registration

Stem Shovel ran the app's own chordsPerBar over the bench's stored Study 1 notes, reproduced the committed 424 of 440 exactly, and found two faults: a bass pickup into the next section renamed the bar it ended (the lowest note earned the root bonus whatever its length), and a power chord's ninth harmonic leaked a faint second over the old 5 percent presence line, so sus2 explained the bar by exactly the leak. v0.231.0 shipped the fixes: a bass note must sound for 15 percent of the bar, and a template tone under 12 percent of the profile counts as missing at a 0.12 cost. The synthetic verdict was clean, 424 to 439 of 440 with the power-chord song going 20 to 28 of 28. The open question was real audio: the app could only approximate it by rebuilding notes from Study 13's stored text lines, losing short notes and all start times, and that reconstruction read the presence rule as a wash or a small loss there. Its brief, committed verbatim in the bench repo, left the proper measurement to the bench and proposed the decision rule itself. The bench registered one parameterized port of the v0.231 matcher, pinned bar for bar against the retired port across all 440 fixture bars at the old constants, and measured the two rules separately, because a joint number would hide one rule behind the other.

Four Gates, Four Passes

The anchor reproduced Study 13 arm 1 exactly for the third time, 29 of 76 and 24 of 86 on freshly re-transcribed notes, so every comparison sits on the stored footing rather than the approximation. The fixture table reproduced the handoff's five configurations to the digit through the bench's registered port, and the fixture gained matcher3 fields beside the v0.141 and v0.142 ones, nothing rewritten. On real audio the bass rule alone lifts the gated song from 29 to 39 of 76 and the exploratory song from 24 to 26 of 86, inside the registered central estimates. And the shipped pair reads 41 of 76 and 30 of 86, a net of plus six over bass-only across both songs, so the presence rule passes the author's own registered decision rule, keep it if and only if its real net is at least zero.

The Overturned Hedge

The satisfying result is whose prediction failed. The handoff hedged its own presence rule, reading it from reconstructed notes as trading strict power-chord wins for majors with a weak third, and the bench registered that hedge as its prediction: a small negative net and a revert recommendation. The proper measurement overturned both. Strict-first accuracy rises too, 20 to 24 on the gated song and 18 to 23 on the exploratory one, so the gain is genuine readings rather than allowance credit. The lossy reconstruction was the artifact, not the rule, and the app's stated judgment, that an under-claimed power chord is the better failure for a chart hint, turned out right on the numbers as well as the merits. The disclosed sweep says the shipped constants sit on a plateau: a bass line of 0.10 readmits the pickups, 0.25 regresses the faster-bass exploratory song exactly as the handoff reasoned, and nothing in the grid earns a re-tune.

The Floor Moves, the Ceiling Does Not

Two constants and one duration check take the real chord path from 38.2 to 53.9 percent on the gated song, which is exactly Study 13's reconciled 41 of 76: the free code change is worth as much on this song as the language-model reconciler pass. The ordering from Study 17 is unchanged and the report says so plainly: ChordMini on the mix still reads 59 of 76, eighteen bars past the improved matcher, so the front-end swap remains the measured next step. v0.231.0 raises the floor the app already ships. The hint baseline for anything measured after it moves from 29 to 41 of 76, and future studies gate their anchors against the version they pin. The verdict delivered: ship stands, nothing to change.