music/bench/ear_xlsft_lm/diffs/MPX_CH02_EXPLORE_r2_XLSFT_LM.diff.json
{
"sample_id": "MPX_CH02_EXPLORE",
"revision": 2,
"target_card": "UNMATCHED__10_-_Desolation_Hong_Kong_canal_",
"generated_card": "NOCORPUS__MPX_CH02_EXPLORE_r2",
"job_bpm": 144,
"scalars": [
{
"dimension": "total_length_s",
"target": 88.346,
"generated": 88.0,
"error": 0.0039,
"scale": "rel:10.0"
},
{
"dimension": "section_count",
"target": 8,
"generated": 11,
"error": 0.375,
"scale": "rel:3.0"
},
{
"dimension": "global_bpm",
"target": 143.55,
"generated": 123.05,
"error": 0.1428,
"scale": "rel:40.0"
},
{
"dimension": "time_to_hook_s",
"target": 0.139,
"generated": 32.264,
"error": 1.0,
"scale": "abs:20.0"
},
{
"dimension": "time_to_full_texture_s",
"target": 22.105,
"generated": 0.116,
"error": 0.8796,
"scale": "abs:25.0"
},
{
"dimension": "flow_peak_position",
"target": 0.7179,
"generated": 0.4761,
"error": 0.2418,
"scale": "abs:1.0"
},
{
"dimension": "mean_lanes",
"target": 3.686,
"generated": 6.592,
"error": 0.7884,
"scale": "rel:2.0"
},
{
"dimension": "max_lanes",
"target": 9,
"generated": 9,
"error": 0.0,
"scale": "rel:3.0"
},
{
"dimension": "silence_fraction",
"target": 0.0618,
"generated": 0.0,
"error": 0.618,
"scale": "abs:0.1"
},
{
"dimension": "drops_per_min",
"target": 20.3744,
"generated": 22.5,
"error": 0.1043,
"scale": "rel:4.0"
},
{
"dimension": "raises_per_min",
"target": 14.9413,
"generated": 11.5909,
"error": 0.2242,
"scale": "rel:4.0"
},
{
"dimension": "breaks_per_min",
"target": 1.3583,
"generated": 0.0,
"error": 1.0,
"scale": "rel:1.0"
},
{
"dimension": "tuttis_per_min",
"target": 1.3583,
"generated": 0.0,
"error": 0.4528,
"scale": "rel:3.0"
},
{
"dimension": "complexity_steps_per_min",
"target": 19.0161,
"generated": 4.0909,
"error": 0.7849,
"scale": "rel:6.0"
},
{
"dimension": "boundary_step_share",
"target": 0.321,
"generated": 0.333,
"error": 0.012,
"scale": "abs:1.0"
},
{
"dimension": "dynamic_range_db",
"target": 35.84,
"generated": 14.31,
"error": 0.6007,
"scale": "rel:6.0"
},
{
"dimension": "dynamics_peak_position",
"target": 0.0,
"generated": 1.0,
"error": 1.0,
"scale": "abs:1.0"
},
{
"dimension": "note_rate_per_s",
"target": 0.894,
"generated": 0.25,
"error": 0.4293,
"scale": "rel:1.5"
},
{
"dimension": "melodic_salience",
"target": 0.8825,
"generated": 0.0,
"error": 0.8825,
"scale": "abs:1.0"
},
{
"dimension": "spectral_centroid_hz",
"target": 629.7,
"generated": 1430.3,
"error": 0.6672,
"scale": "rel:1200.0"
},
{
"dimension": "onset_type",
"target": "fade_in",
"generated": "cold_statement",
"error": 1.0,
"scale": "cat:1.0"
}
],
"curves": {
"flow_curve": {
"available": true,
"rmse": 0.2543,
"target_mean": 0.3114,
"generated_mean": 0.4506,
"note": "energy = f(density, simultaneity, register); each already normalised to its own track's maximum, so this compares SHAPE, not loudness"
},
"simultaneity_curve": {
"available": true,
"rmse": 0.4294,
"target_mean": 0.3515,
"generated_mean": 0.6328,
"note": "lanes sounding at time t, divided by the 10-band ceiling"
},
"tempo_curve": {
"available": true,
"rmse": 0.2824,
"target_mean": 0.6452,
"generated_mean": 0.5886,
"note": "the tempo FUNCTION of section 11.3, divided by a 240 bpm ceiling"
},
"dynamics_arc": {
"available": true,
"rmse": 0.2673,
"target_mean": -0.4349,
"generated_mean": -0.2708,
"note": "the dB arc, divided by a 60 dB window"
},
"complexity_curve": {
"available": true,
"rmse": 0.477,
"target_mean": 0.4879,
"generated_mean": 0.8944,
"note": "the per-bar complexity index of section 11.6, each normalised to its own track's maximum"
},
"register_balance": {
"available": true,
"l1_distance": 0.5098,
"note": "total absolute difference across the 10 octave bands; 0 = identical spectral weighting, 2 = disjoint"
}
},
"scalar_distance": 0.5337,
"curve_distance": 0.3276,
"distance": 0.4512,
"distance_definition": "0.6 * mean scalar error + 0.4 * mean curve error. Each scalar error is the absolute miss divided by its own scale and clipped at 1, so no single runaway dimension can dominate; each curve error is an RMSE on already-unit-normalised curves. 0 is identical grammar, 1 is maximally different on every axis. It is a DISTANCE, not a quality score -- it says nothing about whether either track is good."
}