Study 5 · Composition
The composed pipeline: the chain adds nothing of its own, and its first run audited its own predecessors
Composition
Study Overview
The Question
Studies 3 and 4 measured chart-read and the song-check separately, and separately they are solved problems: a frontier reader at 98.8% and a listener that is perfect when the chart is in view. But production composes them, and the chart beside the mix is chart-read's imperfect output, not the truth Study 4 supplied. Study 5 measured the composed chain: Study 4's 12 metered mixes, their matching lead sheets engraved by Study 3's committed pipeline, read by production's chart-read in both input formats, and the reader's output fed to production's song-check in place of truth, with Study 4's frozen chart-arm cells as the registered ceiling and a committed propagation classifier sorting every fact into pass, listener-added, propagated, rescued or mutated.
The Study That Audited Its Own Predecessors
The first run's biggest result was upstream of the study: the PDF-arm reader kept reporting chords with no flats, and the mechanism check proved the bench's own PDF generation was the culprit, pdfkit's WinAnsi fonts silently dropping the music accidental glyphs Verovio writes into chord text, in every committed PDF of this study and Study 3 alike. The composed chain, built to detect error propagation, caught a fixture confound its predecessor had published numbers on. Amendment 1 committed the finding, the affected cells were discarded and archived, the PDFs re-rendered with ASCII accidentals, and Study 3's Correction was issued the same day. One classifier disclosure rode along: a reader's positional error converts downstream into a confidently wrong second-placement that the value-level extraction attributes to the listener, so the analysis reports position propagation explicitly.
The Corrected Verdict
On honest pages, both gates pass with zero listener-added errors. Across 38 facts per arm the reader made exactly one error, omitting one score's relative-major modulation line from its summary, and the listener rescued it in both arms: reading the chart's chord content (the progression turns to C material), it reported the modulation anyway and placed it at 20.2 seconds against a 20.21-second truth. That rescue is the composed chain's one addition over its parts, and it runs in the right direction. Nothing propagated, nothing mutated: 37 pass and 1 rescued, per arm, against a frozen ceiling the chain equals.
What the Confounded Run Showed
The discarded first run earns its own record, because it demonstrated the failure mode the corrected run no longer could: the chain faithfully transmitting upstream damage. Shown a B-minor sheet whose F#7 printed as F7, the reader reasonably named D major and the listener propagated it one to one; the reader's positional misreadings converted into confidently wrong second-placements downstream. The chain did not amplify or dampen any of it. The corrected run shows none of these, for the simple reason that the pages stopped lying.
The Composition Law
Both runs together give the honest composition story: the chain adds nothing of its own. With a competent reader on honest pages it equals its ceiling; fed corrupted pages it transmits the corruption at full confidence. The quality of the composed feature is exactly the quality of the page its reader is shown, which makes the reader's input, not the chain's plumbing, the thing worth engineering.
What This Licenses
For Stem Shovel: the song-check inherits chart-read's accuracy one to one, so surfacing the chart summary for user correction protects everything downstream of it, and the production model's unprompted key-vs-chords inconsistency notes deserve treatment as a first-class signal rather than prose, since they are the one detectable trace of unrecoverable upstream corruption. Total spend about $3.40 of the $6 fence, both runs included, with the confounded cells archived rather than deleted.