data seat · read-only
melchordgen is a music-generation engine trained on one producer's own catalog, being built into a VST that listens to a sample and generates chords and melodies in his taste. These are the project's daily engineering logs: what was built, what the producer's ratings said, and what changed as a result.
Where it stands: past the 10-fair-test campaign minimum, the fire rate is interpretable: the 17-test era is closed: 8 of 17 samples pass the regen bar (95% interval on the rate: [0.230, 0.722], certified; point estimate 0.47 against the two-in-three target; takes per sample are not fully independent, so intervals read wider in truth). The target is not met. The honest color: every miss carries a measured mechanism, seven lever candidates died at preregistered gates, and the era's second half out-fired its first by 33 points. The rater's own blind calibration tripwire fired at era close and was decomposed by pre-committed rule: his rating noise is zero, his bar rises about a point per week on old material (by design, now measured), and the scorecard is drift-immune because every counted test was rated same-day. The next arc is the producer's call. Single-rater personalization study; the instrument's limits stated here.
Era two passed its measurement threshold: 12 fair tests complete on the fresh sealed pool, and the rate is interpretable: 1 of 12, with the last certified 95% interval at ten tests reading [0.003, 0.445]. That interval overlaps era one's [0.230, 0.722], so the eras are not statistically separable at this depth; the stricter judge (his bar rises about a point a week, measured) and the harder pool ride as context, with no claim made in either direction. The era also produced its first adoption: the rolled-strikes variety mode, three-for-three over its siblings across three grounds (the era's only 4s), adopted on his ordinary ratings alone. Every miss carries a measured mechanism, timing has been silent four straight serves since the window fix, and the front remaining complaint (melodies stopping after bar four) traced to a training truncation cause with its first fix killed at a frozen gate and its direction confirmed. The full serve-by-serve audit is in report 12; report 13 carries the threshold day.
What the conviction actually rests on (per the external review, D214): not the fires, which are statistically thin, but three cases where the project tried to disprove itself and failed: the b28 moment (the highest taste-metric score ever produced was rated down by the ear, proving the ground truth is independent of the machinery); the 0-of-6 scorer kill (a metric adopted in the morning was backtested against his historical verdicts and killed the same day); and the 104-bpm byte-patch (identical music went from a 2 to 4/4/4 when only its tempo stamp was corrected, a one-variable ablation).
compiled by the project's data seat (a build agent) · verdicts remain the producer's · the map (macro view) · methodology & glossary · decisions ledger · batches index · live scoreboard · times US Eastern