Study 10 · Listening
Neural MIR baselines: the classical detector holds its ground, and the home-field rematch cuts both ways
Listening, the detector thread closed
Study Overview
Drafted From the Landscape, Verified by the Bench
The third cross-agent study: the Stem Shovel agent drafted it from its own landscape review, asking whether the app hears with obsolete code, and the bench finished and registered it after verifying that the three cited HuggingFace ONNX models exist with hashes matching their published tables. The candidates: Beat This! small (MIT) for tempo and meter, LinkSeg 7-class (CC-BY 4.0) for sections, ChordMini (MIT) for chords, plus two literature key-profile swaps; one published tracker was omitted on licensing (its library is CC BY-NC-SA). Runners are vendored reference code; the fixture adds a 12-item hard set (swing, octave-boundary tempi, reverb, dense mixes, mediant modulations) and 4 extended-chord progressions. Zero LLM spend in both arms.
Five Gates, Five Fails, All in Production's Favor
The synthetic arm's verdict is uniform: every registered gate failed in the shipped code's favor. The beat gate carries the study's most instructive nuance: at the registered plus-or-minus 1 bpm, production is exact on 29 of 33 to the neural tracker's 18, but the tracker derives tempo from a 20 ms frame grid, so it answers 142.9 for a true 140, musically right and numerically outside the letter; a disclosed sensitivity read at plus-or-minus 3 is a dead tie at 29 each, the program's fifth registration lesson (an absolute tolerance at DSP precision measures the candidate's output quantization, not its hearing). What the letter cannot take away: the tracker's per-beat grid scores an F-measure of 0.892, something production does not have at all, and it answered the pulseless pads honestly. Meter went 27/29 to 21/29 for production, with the hoped-for free 6/8 dead at 1 of 4; and the key profiles tied on synthetic material (29, 29, 30 of 33), deferring the real question to the amendment.
The Segmenter and the Chord Reader
The section gate failed decisively: LinkSeg placed 54 false boundaries across the 22 single-section items, about 2.5 per item, worse than Study 8's chart-anchored language models, while finding only 5 of 8 contrast boundaries; graph attention does not cure the loop-invention disease on this fixture, where Study 9's free activity detector scored 8 of 8 with zero. The chord gate split: the production matcher took the full vocabulary 483 to 418 of 504 bars (the neural model hears triads where the truth says sevenths), ChordMini edged the hard set on major-minor nuance 61 to 59 of 64, and one collapse earns its name: the N.C.-bar song fell to 4 of 28 under the neural reader while the matcher read 27. Deployment was never the blocker; every candidate fit the size and latency budget, and accuracy decided everything.
The Home-Field Rematch
The registered real-recordings amendment ran the same day on a public catalog of the band's own songs, scored against the band's declarations, including declared section start seconds. The domain warning was real, in both directions. Sections flip outright: LinkSeg finds real form at 29 of 33 declared boundaries within 4 seconds, including one song at 8 of 9 with zero false boundaries, while Study 9's D-activity, perfect on synthetic arrangements, drops to 25 of 33 with 19 false boundaries on real stems. That is a plain correction of Study 9's ship suggestion, and this page says so: the measured no-chart section fallback is now the 1.5 MB CC-BY segmenter, not the bench's own detector. Tempo kept its known shape with real teeth: production's two misses are gross, an exact half of a declared 224 and an exact two-thirds of a declared 145, the 3:2 family on real recordings, while the neural tracker leads 3 of 4 at the wider read (n of 4, direction not proof) by degrading gracefully instead of octave-jumping. Meter held for production, which read the real waltz the tracker missed. And the key question closed hard: Krumhansl 4 of 6 against 1 of 6 for both literature profiles, which collapse into parallel-minor and subdominant flips on real mixes. The synthetic arm said interchangeable; the real arm says production's choice is the right one.
Ship Guidance
Briefed back to Stem Shovel: nothing ships for tempo, since the detector's failure classes are known and the neural alternative is not plus-or-minus-1 accurate on real material either; keep Krumhansl without a second thought; and consider LinkSeg, behind the chart when one exists, as the no-chart section fallback that the real recordings earned it. The honest boundary runs both ways and is stated in the report: the synthetic corpus is the material production's templates were born on, and the real catalog is the material the neural models were trained for; both arms exist precisely so neither home field decides alone.
The Scorecard, and the Ledger
The registered scorecard ran against the draft's hopes: prediction 3 right in full including the one-profile detail, 1 wrong at the registered letter and right at the disclosed wider read, 2 wrong in both clauses, 4 wrong except the hard-set edge, 5 half-right with the zero-false clause failing much harder than predicted, 6 wrong in the happy direction. The ledger: zero dollars in both arms, about 30 MB of hash-pinned weights, an 800 MB public-audio fetch with hashes committed, and the detector thread that began with Study 4 now closed end to end: the models read, the detectors hear, and every shipped choice is a measured one. The octave classes this study left standing were then retired by representation in Study 11's tempo ladder.