Batches index
Every rating batch since the his-data era began: what it tested, what the producer's verdict was, and the
report where it happened. Ratings are 1-5; see the methodology page.
Batches before b25 belonged to the earlier rule-based era and are summarized in the project ledger.
the his-data era · report 01
b25Jul 6First A/B of a melody model trained on his own clips vs the old baseline.3 vs 2 · his-data wins
b26Jul 6Rule-generated chord arps on a new sample.all 2s · rules retired
b27Jul 6Chords generated from his own chord clips (F-minor sample).4 · his best-chords verdict yet
the climb · report 02
b28Jul 7Voicings built to match his statistical fingerprint (him-score 0.91).3/2/2 · lost to the ear
b29Jul 7Cohesive generation: melody conditioned on the chords (the connections model).all 4s once timing-grounded
b30Jul 7Melody density: denser fill vs sparse.4 · denser wins
b31Jul 7Generate-many, select on melodic interest.first 5 (moderate, not max)
b32Jul 7The recipe on a D-major sample.all 3s · major color gap
b33Jul 7Harmony adapter from his liked tracks added (the color layer).5 · color layer adopted
b34Jul 7Forced 33% out-of-key "color" (the wrong proxy, tested directly).all 1s · L19 born
b35Jul 7In-key character; flagged the plain dominant chord in major.3/3/4
b36Jul 7V-to-vi substitution + melody conditioned on chords (wrong-notes fix).4/4/4 · substitution + conditioning adopted
b37Jul 7Progression variety in major.4/3/4 · initiation lever named
b38Jul 7Four chord-timing feels tested.3/3/4/4 · call-response wins
b39Jul 7Call-response refined + clash fix + more melody movement.4/5/5 · call-response adopted
b40Jul 7The full stack: likes color at 120 tracks + loved-melody blend.5 · static-hold reclassed as bridge material
b41Jul 7The same recipe on F-minor (generalization).3/4/4 · travels
b42Jul 7Three F-minor progressions (a chord vote).4/4/2 · L20 born
b43Jul 7Melody A/B: sweet-spot-trained models + densest take vs the old path.5/4 vs 4 · new default
b44Jul 7Third-chord options on the winning progression, b9 guard applied.5/5/4 · off-kilter named
the verdict day · report 03
b45Jul 8Stacks-vs-assembled A/B scaffold; the stacks model was banked before any clip earned delivery.never rated
b46Jul 8First auto-screened library keepers (no ear in the selection loop).5/5, re-confirmed 5/5
b47Jul 8Blind A/B: the engine trained on the duplicated corpus vs the deduplicated rebuild.4/4 tie · clean adopted
the lever stack · report 04
b48Jul 9Three fresh samples, two never-used keys. v1 failed on windows and tuning; v2 on sample audibility.2/1/2, 2/2/2 · parked, robustness layer built
b49Jul 9Blind selector A/B: current ranking vs context-adaptive band-fit.4/4 tie · adaptive adopted
b50Jul 9Blind voicing A/B: rule-builder vs retrieval from his own chord vocabulary.3 vs 1 · rules keep it; L21 born
b51Jul 9Blind chord-initiation A/B: fixed grid vs adaptive (his b37 lever, built to measured bands).5 vs 4 · the bounce win; L22
b52Jul 9Blind progression-source A/B: the likes-trained model vs his-corpus retrieval.4/4 tie · model default; retrieval = off-kilter cell
b53Jul 9Blend depth round 1: standard 1.00 vs the lone 1.30 survivor (D-major).5 vs 4 · deep gem real
b54Jul 9Blend depth round 2: deep-mined 1.30 (168 takes, 2 survivors) vs 1.00 on F-minor.3 vs 4 · standard keeps default
b55Jul 9Blend depth round 3: standard 1.00 vs light 0.85 (fresh progression).3 vs 4 · light stays internal
the measurement era · report 05
exp 1Jul 10Held-out exposure: crystalpeak, A minor, a key the engine had never generated in.0/3 (2/2/3) · chord-on-chord found
exp 2Jul 10Held-out exposure: MELODYtool, G minor; blind validation of the chord-following fix.0/3 (3/3/3) · correct but static
exp 3Jul 10Held-out exposure: MissYou, C-sharp major; the free-gen path on fresh static ground.0/3 (2/1/3) · too changey; L23
exp 4Jul 10Held-out exposure: Gun Barrel, C-sharp minor sounding; blind validation of the character-driven chord-rate dial (gate read the sample as a drone; two chords over the window, preregistered).0/3 (3/2/2) · chords held, timing named; L24
b56Jul 10Interlude timing gate, blind, zero pool cost: one shared sparse melody on trusty D-major; adaptive gap-answers vs metric strong-beat strikes.4/4 tie · metric non-inferior; exposures go metric
t2 1Jul 10Home-field exposure (same-source, never counted toward the milestone): the kart chord stem, F-sharp minor, 102bpm. First borderline routing (0.573, free-gen) shipped as decided, wrong-routing prediction on record.0/3 (2/3/2) · prediction confirmed; threshold 0.60→0.50; self-check fired
t2 2Jul 10The kart-stem re-run, both tier2-01 lessons aboard: follow at 0.50 (collisions 18-20 to zero) + grid-establishing strikes.2/3 FIRE · era's first fires; the in-time qualifier exposed the cut bug
t2 3Jul 10The kart stem regenerated on the corrected downbeat-anchored cut (intake v2), standard strikes; both readings preregistered.2 · no fires; confounded; his tempo question cracked Entry 9
t2 4Jul 10tier2-02's exact fire clips byte-patched 104 to 102.000 bpm, zero regeneration: the pure clock test.4/4/4 · 3/3 FIRE, first full sweep; the timing era closes
exp 5Jul 10Held-out exposure, milestone lane: the 135bpm loop, 135bpm, never seen; the first fully clean pipeline (true clock, purity cut, follow stack, variety axis, scripted 9-check precheck).5/5/5 · FIRE 3/3, the milestone's first fire-datapoint
exp 6Jul 10Held-out exposure, milestone lane: EH, a true drone in B minor; variety axis on free-gen ground (three distinct starting chords); the precheck's first pre-ears save.5/5/3 · FIRE 2/3, datapoint two, consecutive
exp 7Jul 10Held-out exposure, milestone lane: the clean era's first pulsed ground (real drums, G-sharp major, 120bpm); span-v2 and variety starts aboard.3/3/2 · no fire, one take short; density axis exposed
exp 8Jul 11Held-out exposure, milestone lane: Elevator, the last original-pool loop; rhythm feels A/B/C debut, retro roll debut.2/2/2 · sample FAILS the bar; repitch confirmed +2 semitones same day
beautify-01Jul 11Re-listen lane (never counted toward the bar): the dense-wash ground rebuilt with the thinning fix; roll-trim debut.3/unscored/3 · no fire; complaint moved to register, retune + guard shipped same evening
exp 9Jul 11Held-out exposure, milestone lane: PROTON (D-sharp minor, new artist), the refill's first draw; repitch clause and register guard debut armed.4/4/4 · sample PASSES; thin by the producer's own note; board 3/5 provisional [0.147, 0.947]
beautify-02Jul 11Re-listen lane (never counted): the repitched Elevator ground re-served with the follow-the-shift lever, one lever changed.5/5/4 · zero notes where the morning read 2/2/2; banks nothing, fresh repitch ground decides
exp 10Jul 12Held-out exposure, milestone lane: alez, first counted serve from the pruned tree; follow route; key-confidence flag preregistered both ways.3/3/3 · sample FAILS; suspect cleared by his ear; struck-half-step screen class found, fixed, backtested same hour
exp 11Jul 12Held-out exposure, milestone lane: ELIJAHKING, first serve with the sting fix live; the repitch detector's closest borderline call (0.833, held fire).4/4/4 · sample PASSES, no qualifier; the borderline call's adaptation praised at the ear; board 4/7 [0.184, 0.901]
exp 12Jul 12Held-out exposure, milestone lane: dinuovo, the densest ground served; first reference score (the ground's own transcription rated after the takes).3/3/3 · sample FAILS, zero complaint notes; reference scored 3 = the ceiling was the material; wall class ruled library-hard; board 4/8 [0.157, 0.843]
exp 13Jul 12Held-out exposure, milestone lane: AUG, major-key ground; the era's closest routing call (0.471) and the first post-takes reference serve.5/4/4 · sample PASSES 3/3; the era's first fresh 5 and the first-ever major-key fire; board 5/9 [0.212, 0.863]
exp 14Jul 13Held-out exposure, milestone lane: solofulo, overnight batch one; follow route at 0.619, low key-confidence band flagged both ways.5/4/5 · sample PASSES with two 5s; the 10-fair-test minimum reached, the rate reads 6/10 [0.262, 0.878], earned benchmark not met
exp 15Jul 13Held-out exposure, milestone lane: OPTOMO, overnight batch two; the disclosed cut-only tuning correction's first ear test.3/4/3 · sample FAILS (1/3); mood-fit named (bright against a dark ground, in key); rate 6/11 [0.234, 0.833]
exp 16Jul 13Held-out exposure, milestone lane: PASSIONE, the Elevator-tell ground; repitch trigger never armed (not the banking test).3/3/3 · sample FAILS; the arp class named (taxonomy's fourth); staleness meta-complaint measured to two nulls; rate 6/12 [0.211, 0.789]
exp 17Jul 13Held-out exposure, milestone lane: TRY-AGAIN, the quota batch; maximal harmonic movement, the tempo ambiguity preregistered both ways.4/2/3 · sample FAILS (1/3); weak loop with no excuse granted; the spice direction note flagged for his ratify; rate 6/13 [0.192, 0.749]
exp 18Jul 14Held-out exposure, milestone lane: HUB, the messy tail's first draw; F-minor home ground, the +49-cent cut-only correction disclosed as first suspect.4/3/4 · sample PASSES; board 7/14 [0.230, 0.770]; correction class validated n=2 silent; home-key leans agnostic
exp 19Jul 14Held-out exposure, milestone lane: MOSQUITO, tail draw two; the 0.603 key-confidence floor preregistered as first re-read.1/1/2 · sample FAILS at the era floor; the section-choice idea tested and killed same-hour at its prereg gate; hard-song finding; board 7/15 [0.213, 0.734]
exp 20Jul 14Held-out exposure, milestone lane: aug-obsession, tail draw three; the repitch clause silent by condition order.3/2/4 · sample FAILS (1/3); the twice-flagged chord's mechanism measured (ambiguous-cluster root pick) before any change; board 7/16 [0.198, 0.701]
exp 21Jul 14Held-out exposure, milestone lane: bumiao, the tail's last draw; three standing quality flags including the correction class's largest test (+39c).4/4/4 · sample PASSES; THE ERA CLOSES at 8/17 [0.230, 0.722], the 2/3 target not met; every miss mechanism-measured
serve 02Jul 14Calibration lane (truly blind re-serve, unsealed post-verdict): the rater re-tested against his own week-old ratings.deltas -1/-1/-1 · the kill line fired by its own rule (kappa 0.571); ruled: noise zero, drift ~1/week = the rising bar measured; scorecard drift-immune by construction