Study 13 · Listening
The chord path meets real audio, and the reconciler earns its keep
Listening, the conditional
Study Overview
The Conditional Registration
Would a chord reconciler still have value? Study 13 answers it the way the program has learned to ask: as a registered two-arm conditional, measuring the thing the reconciler would improve first and running the reconciler only if the measurement leaves headroom, with the trigger registered at gated accuracy under 90 percent. Study 12's artifacts made the measurement possible at all, supplying the bar grid in recording seconds and bar-aligned chord truth on real audio. Arm 1 cost nothing; the whole study ran at about $0.60 of its $3 fence.
The Home-Field Gap, Measured
Arm 1 is the finding that matters most: production's chord path, Basic Pitch plus the v0.142 matcher, which scores 95.8 percent at home on the synthetic bench and beat the neural reader in Study 10, reads the real letter-lands recording at 38.2 percent, 29 of 76 bars, with the registered power-chord allowance doing real work on nine of them, and the exploratory song at 27.9 percent. The predicted home-field gap was 15 to 30 points; the measured gap is nearly 60. The binding constraint is the transcription on real timbres, the rich-harmonics weakness Study 1's fixture work warned about, now measured in production conditions instead of suspected. The trigger fired without drama.
Amendment 1, the Recurring Own-Harness Lesson
All four pilot cells truncated at the registered 4000-token output cap, because the models narrate their way through every bar before emitting the JSON: the exact budget-conflates-with-failure class Study 8's amendment caught, back in the bench's own harness. The cap moved to 16,000, the cells re-ran under the fence, and the truncated run stays in history. One more disclosure rode along: the production model ignores the registered temperature-zero deviation, which the gateway reports as unsupported.
The Reconciler Earns Its Keep
Arm 2 passed its gate for both models. One text-only pass over the key, the detected chords and production's own per-bar note lines took the production chart-draft model from 29 to 41 of 76 bars, 53.9 percent: twelve fixes, zero breaks, across 27 changed bars, so the double-fix gate registered against Study 9's agreement trap (fixes must at least double breaks) passed at infinity. gemini-3.5-flash climbed from 29 to 36 with seven fixes and zero breaks. On the exploratory song the production model added three more bars without breaking any, while gemini burned the full 16,000-token budget narrating 86 bars and never reached the JSON, the budget failure mode of that family from Studies 2 and 8, not the over-editing mode the prediction expected.
Ship Guidance
Two findings, delivered in order of weight. First, chord drafts on real user audio are far weaker than any synthetic number suggested, and that reframes the feature's honest quality bar: a UI hedge is worth shipping now, regardless of anything else. Second, the reconciler is real and cheap, a text-only pass by the model production already runs for chart drafts, lifting the real-audio number 16 points with zero breaks for about a cent a song: ship it inside the chart-draft job, symbols in and symbols out, keep the detector's chord whenever the model's answer parses to nothing, and cap output generously. The honest ceiling stands: 54 percent reconciled is better, not good, and if the chord feature matters, the next investment is a stronger transcription on real timbres, not more model.
The Scorecard, and the Arc Note
The registered scorecard: 1 right in direction and wrong in magnitude; 2 right, with the changed-bars count just over its sub-claim; 3 wrong in the instructive direction, gemini failing by budget rather than by over-editing; 4 right. The disclosure is Study 1's, unchanged: the analyst is Claude Fable 5 and arm 2 measures claude-fable-5-1, tolerable because every grade is mechanical and committed before any call. And the arc note this page exists to carry: this is the program's first study in which a language model IMPROVED a hearing pipeline, and it did so by never hearing anything, symbols in and symbols out, which is the twelve-study law applied rather than broken.