03worked / didn't
The honest board: adopted against killed
Everything below entered through the same door: preregistered criterion, then evidence, then a
ruling. The adopted column holds because each line traces to a blind verdict; the killed column is
the record's proudest asset, because every entry died at a gate written before the data came back.
Adopted, verdict-traced
ADOPTEDCohesive generation - melody conditioned on chords, made together (b29/b36; fixed the recurring wrong-note class)
ADOPTEDClose voicing + in-key color - beat the statistically "most-him" wide voicing at the ear (b28 vs b29; the original ear-beats-metric case)
ADOPTEDGenerate-many, select at moderate - maximum-interest selection overshot to "too emotional" (b31)
ADOPTEDAdaptive chord initiation - chords answer the melody in its gaps; the lever he named himself, built to bands measured from his likes (b37 → b51, 5 vs 4)
ADOPTEDThe follow route - ride a moving sample's harmony as the map, voice fresh (exposure-1 miss → threshold ear-calibrated 0.60 → 0.50; strongest ear evidence at sample 10)
ADOPTEDThe character dial - the sample's own character drives harmonic rhythm (three misses unified into one law, L23; held blind on its first test)
ADOPTEDThe screen gauntlet - clash, leap, floors, tense-hold, naked-sting, register ceiling, ending hang; each minted by a rating, several backtested against every loved clip before shipping (0 false trips on 212)
ADOPTEDReference score + certified intervals + preregistration - the measurement culture itself: ground quality measured per serve, no number printed uncertified, no outcome spinnable
PROVISIONALRepitch-follow route - built from a measured +2-semitone mechanism, held 5/5/4 on its calibration ground (banks nothing); awaits a fresh repitch ground
PROVISIONALDensity-thin + retuned lift window - the wash-class fix; its complaint class did not recur, its window retune ear-validates on the next dense fresh ground
PROVISIONALAdvisory deep-key flag - a second detector adopted at exactly its measured scope (pitch-set only, flag-never-override); corroborated his ear on both contested grounds
Killed at preregistered gates
KILLEDThe him-score as a target - the 0.91 DNA-match rated "odd" (b28); metrics have been selectors, never synthesis targets, ever since
KILLEDBorrowed-chord percentage - out-of-key color rated 1s across the board (b34); color means in-key richness
KILLEDEmbedding take-ranking - adopted in the morning, backtested against his historical verdicts by evening: 0 of 6 agreement, dead the same day
KILLEDRetrieval voicing - his own chord vocabulary, retrieved by frequency, lost 3 to 1 at the ear (b50); transitions are the unit of taste, not shapes
KILLEDRaw sting-count screen - died at its own backtest: his loved clips carry more raw semitone pairs than the batch he flagged; only the naked-sting anatomy separates love from pain
KILLEDMelody-continuation lever - refuted by its own probe (the fires decay identically; the decay is his own phrasing); re-opened once at its preregistered door, measured inconclusive, closed again at its gate
KILLEDMood-fit axis - real complaint, era-wide non-predictor (9-point gap under a 15-point preregistered gate); the measure stays in the toolbox
KILLEDStaleness lever - the premise inverted twice at the bench: high-repetition serves fire more, and the era's second half out-fired its first by 33 points (the improvement curve, measured)
KILLEDSection-choice hypothesis - does the intake pick the wrong part of multi-section songs? Instrumented and answered no, same hour, criteria sealed before looking
KILLEDSync-aware selection - the profile's largest measured gap (21/21 batches under-syncopated), converted to a selection term and rejected 4-1 blind at his ear; corpus-closeness does not translate to felt pocket; the gap stays measured, awaiting a different mechanism
BLOCKEDThe transposition-augmented retrain - the research literature's top recommendation, tested blind: lost 7-1 at his ear and failed to generate on 2 of 8 grounds; the incumbent holds
BLOCKEDThe chord-model recook - a fully-documented rebuild, blocked at the sealed line twice by his blind ear; the reject is the finding: the fossil carries preferred behavior a near-identical recipe did not recover
BANKEDWhole-loop stacks - three cooks hit the same data ceiling (3,086 unique sections); banked on its own pre-written kill condition, unlocks on corpus growth
BANKEDPocket-matching microtiming - designed, waiting for its first "feels stiff" ear evidence; the literature itself calls the fingerprint contested
05the ground map
The ground-class taxonomy: where the engine wins and where the world is hard
PASSESClean moving grounds - the follow route's home turf (5/5/5, 4/4/4, 5/4/5 among them); the thesis's strongest evidence
PASSESDrones - grid-establishing chords become the clock (5/5/3); span schedule tuned by his ear
PASSESMajor-key grounds - the early eras' named gap, closed with the campaign's first fresh 5 (5/4/4)
PASSESNear-repitch two-section grounds - the detector's hardest call held fire and his ear praised the resulting adaptation
LEVEREDDense in-key washes - the density-thin route built from the measured redundancy mechanism; complaint class has not recurred
LEVEREDRepitched chord hits - mechanism measured to the semitone, route built and held on calibration ground; awaits fresh proof
HARDWalls (11+ pitch classes sounding) - ruled library-hard by two instruments including his own reference score; draw proposals carry the advisory
HARDHard songs - grounds whose own musical ambiguity caps any complement; the reference score converts these from confusion into measurement
The frontier, in the producer's own words and on his card: consistently
great complements on musical inputs is the standing target; elevating weak grounds is a flagged
someday. The map redraws as verdicts land.
The Map · compiled by the data seat · updated 2026-07-16, era D266 · every number certified before printing · sources: the sealed daily reports 01-10 + the live report, the decisions ledger, the engine manifest