Findings No. 81
Section sensing survives grid failure
Stem Shovel's livery placed all 11 section boundaries correctly even while its harmony timeline was wrong, showing structure and harmony detection fail independently.
Stem Shovel's section sensing and harmony analysis are two separate judgments, and a failure in one does not predict a failure in the other. On one track, livery placed all 11 of 11 section boundaries correctly while its harmony timeline for the same track was wrong. The chord and key grid was off, but the verse, chorus, and bridge boundaries drawn around that grid were not.
The fix was to stop treating one analysis as a referendum on the other. Section detection and harmony detection are reviewed as separate passes, each judged on its own evidence, so a wrong chord grid does not get used to discount a correct section call, and a correct section call does not get used to wave through a bad harmony read.
The general rule: when a system makes more than one kind of judgment about the same input, check each judgment on its own terms. A pipeline that bundles structure and harmony into a single verdict will throw out a good result along with a bad one, and the fix usually costs nothing more than scoring the two separately.
Receipt: Stem Shovel (livery).