Read the song as a whole: section repetition and key plausibility as evidence for a bar's chord

[eolivelli]

Enrico's proposal from the il-mio-canto-libero field pass (2026-08-22), in his words: the only way of tagging the chords a quiet accompaniment leaves ambiguous is "to evaluate the bars from an harmony perspective (FA#7 is unusual in a song in key A) + looking at the song as a whole, and looking for repetitions of sequences" — while acknowledging it is still a very hard call.

This converges with what the measurements already concluded independently: #662's four refuted guards for #583 ended on "whatever closes #583 needs evidence the run does not carry — the key (#549), the form, or a root's runs read against each other." The three pieces:

• Key plausibility is The quality decision is key-blind, so it can name a chord with a note the key does not hold #549 (the quality decision is key-blind). It must stay a prior, not a veto — the borrowed-iv package (A minor chord a recording states once on a root it otherwise plays major is withdrawn with the false ones #583/A benchmark for #583, and why four guards for it do not work #662) exists precisely because the out-of-key chord is sometimes the truth — and it must survive modulation, which The key layer holds one span for a whole piece, so a modulation the chords track cannot be said #441 says the key layer cannot yet express (one span per piece; il-mio-canto-libero itself modulates A→B and MW finds both regions).
• Repetition of sequences is the new piece no issue covers: pop states its progressions many times, and the repeats are independent measurements of the same bar. Concrete field case: a bar where Enrico would put D and MW printed E F#m — if the section's other repeats say D there, the repeats outvote the fold. Nothing in MW today compares a bar to its structural siblings.
• The song as a whole is what both need: some notion of form/section identity, which nothing in the pipeline currently produces.

Related: #583 (the consumer that needs run-external evidence), #549, #441, #458, #668 (the honest fallback where no evidence source reaches).