Study 9 · Listening
Stem-aware analysis: the stems name the form, the sum keeps the time
Listening, the detector angle
Study Overview
Drafted by the App, Run by the Bench
Study 6b was registered by the production agent; Study 9 runs the collaboration the other way, drafted by the Stem Shovel agent and finished, registered, and run by the bench. The question is the app's own: it holds the individual stems of every song, yet its detector and both AI features analyze the summed mix, so does stem-aware analysis beat the sum? Four zero-model-spend DSP instruments ran over 21 items (the 15 pulsed Study 4 and 8 specs plus 6 new misleading-stem specs whose fourth part, a rubato lead, an off-grid pad, a noise and room track, or a double-time comp, is designed to pollute the summed onset envelope), with stems shuffled so no instrument knows which is the drums. D-prod is a verbatim port of production's analysis code, pinned at its commit. A 12-call AI side arm rode along; total spend about $0.08.
The Form Half: a Detector That Cannot Invent
S9-G1 passed perfectly on both the truth bar grid and the production-portable detected grid: D-activity, which marks a boundary where the set of active stems changes and holds for two bars, found 8 of 8 contrast boundaries with zero false boundaries on the uniform loops and the change items alike. Set beside Study 8's results on the same items, the shape is the point: models blind invented 3 to 6 sections per loop, the chart-anchored production model still invented about one, and the activity detector invents nothing because on a loop the active set never changes; the failure mode is excluded by construction, not suppressed by a bound. Its blindness is equally structural: all six tempo and key-change boundaries, where the arrangement does not change, are invisible to it, as registered; those belong to the chart. One honest caveat from the mislead class: a noise track with slow swells toggles its own activity bit (two false boundaries on one item), so a shipped version wants the atonal-stem guard.
The Time Half: the Failure Is the Finding
S9-G2 failed decisively, and the failure is the study's most useful result. Keeping the sharpest single stem's tempo instead of the sum's does not help; it harms: D-prod exact on 18 of 21 items, the sharpest-stem rule on 11 of 21, with 7 regressions (5 of them octave errors, because a clean bass line autocorrelates beautifully at half time) and zero rescues. Meanwhile the summed envelope survived all six misleading extras untouched: pad, noise, rubato, and both double-time comps left it exact. Summing the stems is not a compromise production settled for; the cross-stem reinforcement at the true period IS the tempo signal, and the shipped design is vindicated against every stem-aware variant tried. The sum's only misses are a shared systematic, the 3:2 dotted-pulse confusion on compound meters, which no stem-aware variant fixed either. The registered octave-disagreement flag fired on 10 of 21 items including clean ones, a half-time bass being normal music rather than a hazard signal, and is reported as measured: useless as shipped.
The Key Sting
S9-G3 was unevaluable under its registered fallback, since production's key detection was right on 20 of 21 items and there was nothing to separate. The descriptive warning that rode along is worth the section: the one wrong answer carried a per-stem agreement score of 1.00, the maximum, while production's own correlation margin ranked that same answer third-lowest of 21. On this corpus the proposed agreement score did not merely fail to prove itself; it pointed the wrong way on the only error available, while the existing confidence pointed the right way.
The Pilot, and the Familiar Prior
The 12-call pilot gave gemini-3-flash the isolated bass and comp stem in place of the mix. It helped nowhere and hurt twice, with two items flipping from right to wrong, and the tempo answers were the familiar 120, 126, 128 genre prior in both conditions: Study 4's reading-not-hearing, unchanged by stem isolation. The registered prediction said at most one item would change; two did, both for the worse, direction right and count wrong. The fenced matrix stays open but unneeded, since the gates did not require it.
Ship Guidance, and the Scorecard
Briefed back to Stem Shovel: ship nothing for tempo and key, since the summed-envelope detector beat every stem-aware variant and its one weakness is not fixed by stems; and consider shipping D-activity as a browser-side section detector, zero model cost, structurally immune to the loop invention that survived even chart anchoring in Study 8, guarded against sustained-noise stems, with the model kept for names and the chart for harmonic boundaries. The scorecard: prediction 1 right in full, 2 refuted in both halves, 3 right except the agreement-floor guess (the real floor was indifference to the one actual error), 4 right in direction and wrong in count.