music/bench/ear_xlsft_lm/diffs/MPX_TITLE_MAIN_r2_XLSFT_LM.diff.json
{
"sample_id": "MPX_TITLE_MAIN",
"revision": 2,
"target_card": "EX_110",
"generated_card": "NOCORPUS__MPX_TITLE_MAIN_r2",
"job_bpm": 99,
"scalars": [
{
"dimension": "total_length_s",
"target": 141.584,
"generated": 142.0,
"error": 0.0029,
"scale": "rel:10.0"
},
{
"dimension": "section_count",
"target": 7,
"generated": 18,
"error": 1.0,
"scale": "rel:3.0"
},
{
"dimension": "global_bpm",
"target": 99.38,
"generated": 129.2,
"error": 0.3001,
"scale": "rel:40.0"
},
{
"dimension": "time_to_hook_s",
"target": 0.035,
"generated": 0.325,
"error": 0.0145,
"scale": "abs:20.0"
},
{
"dimension": "time_to_full_texture_s",
"target": 3.808,
"generated": 0.673,
"error": 0.1254,
"scale": "abs:25.0"
},
{
"dimension": "flow_peak_position",
"target": 0.3799,
"generated": 0.407,
"error": 0.0271,
"scale": "abs:1.0"
},
{
"dimension": "mean_lanes",
"target": 7.269,
"generated": 6.087,
"error": 0.1626,
"scale": "rel:2.0"
},
{
"dimension": "max_lanes",
"target": 9,
"generated": 9,
"error": 0.0,
"scale": "rel:3.0"
},
{
"dimension": "silence_fraction",
"target": 0.0307,
"generated": 0.0002,
"error": 0.305,
"scale": "abs:0.1"
},
{
"dimension": "drops_per_min",
"target": 11.0182,
"generated": 19.4366,
"error": 0.764,
"scale": "rel:4.0"
},
{
"dimension": "raises_per_min",
"target": 6.7804,
"generated": 15.2113,
"error": 1.0,
"scale": "rel:4.0"
},
{
"dimension": "breaks_per_min",
"target": 2.5427,
"generated": 0.0,
"error": 1.0,
"scale": "rel:1.0"
},
{
"dimension": "tuttis_per_min",
"target": 12.2895,
"generated": 0.0,
"error": 1.0,
"scale": "rel:3.0"
},
{
"dimension": "complexity_steps_per_min",
"target": 9.3231,
"generated": 5.0704,
"error": 0.4561,
"scale": "rel:6.0"
},
{
"dimension": "boundary_step_share",
"target": 0.182,
"generated": 0.417,
"error": 0.235,
"scale": "abs:1.0"
},
{
"dimension": "dynamic_range_db",
"target": 10.59,
"generated": 14.92,
"error": 0.4089,
"scale": "rel:6.0"
},
{
"dimension": "dynamics_peak_position",
"target": 0.4829,
"generated": 0.0,
"error": 0.4829,
"scale": "abs:1.0"
},
{
"dimension": "note_rate_per_s",
"target": 2.232,
"generated": 0.38,
"error": 0.8297,
"scale": "rel:1.5"
},
{
"dimension": "melodic_salience",
"target": 0.8372,
"generated": 0.3532,
"error": 0.484,
"scale": "abs:1.0"
},
{
"dimension": "spectral_centroid_hz",
"target": 2351.9,
"generated": 1324.9,
"error": 0.4367,
"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": 0.2666,
"target_mean": 0.6131,
"generated_mean": 0.4374,
"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": 0.3065,
"target_mean": 0.7232,
"generated_mean": 0.6061,
"note": "lanes sounding at time t, divided by the 10-band ceiling"
},
"tempo_curve": {
"available": true,
"rmse": 0.3169,
"target_mean": 0.7279,
"generated_mean": 0.5909,
"note": "the tempo FUNCTION of section 11.3, divided by a 240 bpm ceiling"
},
"dynamics_arc": {
"available": true,
"rmse": 0.1798,
"target_mean": -0.1985,
"generated_mean": -0.2713,
"note": "the dB arc, divided by a 60 dB window"
},
"complexity_curve": {
"available": true,
"rmse": 0.2266,
"target_mean": 0.8436,
"generated_mean": 0.8874,
"note": "the per-bar complexity index of section 11.6, each normalised to its own track's maximum"
},
"register_balance": {
"available": true,
"l1_distance": 0.6851,
"note": "total absolute difference across the 10 octave bands; 0 = identical spectral weighting, 2 = disjoint"
}
},
"scalar_distance": 0.4779,
"curve_distance": 0.2732,
"distance": 0.396,
"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."
}