music/bench/ear_xlsft_lm/spec_r2_xlsft.json
{
"_doc": "MEASURE -> GENERATE -> MEASURE -> IMPROVE proof spec. Emitted by harness/music_gen/masterpiece_proof.py; do NOT hand-edit.",
"_legal_posture": "Every number here was MEASURED from audio Josh lawfully owns. No exemplar audio is copied, redistributed, or used as model input of any kind; what crosses is a number and an English clause derived from it.",
"_register": "DRAFT-TIER BEDS. Per the melody-first ruling, ACE-Step realises beds and draft cues; the hummable leitmotif layer stays composer-in-loop and is not claimed here.",
"revision": 2,
"emitted_at": "2026-08-06T13:14:58+00:00",
"required_generator": {
"dit": "acestep-v15-xl-sft",
"lm": "acestep-5Hz-lm-4B",
"provenance": "BENCH ARM B — the documented quality tier, PIPE_MUSIC_MODELS bench 2026-08-06",
"control_surface_evidence": "build/audio/bench/acestep_xlsft_lm/control_surface_xlsft_lm.json"
},
"derived_from": [
{
"sample_id": "MPX_CH02_EXPLORE",
"revision": 2,
"purpose_class": "EXPLORATION_ENVIRONMENT",
"target_card": "UNMATCHED__10_-_Desolation_Hong_Kong_canal_",
"target_title": "Desolation (Hong Kong canal) -- Deus Ex, Alexander Brandon",
"target_why": "The measured corpus's sparsest sustained-tension cue: mean 3.69 simultaneous lanes against a 9-lane ceiling, 6.2% silence budget, 2 tutti arrivals in 88 s, and its energy peak held back to 72% through. That is the grammar of low-and-tense that rises late -- which is what the chapter's own music_mood row asks for.",
"canon": [
{
"file": "docs/spine/CH_02.md",
"line": 234,
"text": "music_mood -- survival-horror-opener mood, low-and-tense rising to the crater-lake climax; instrumentation drawn from the Flores gong-waning tradition at reference register (Thread 21, the first vril-active-instrument demonstration); combat-cue register for the stalk"
},
{
"file": "build/audio/music_cue_table.csv",
"line": 0,
"text": "MC_0026 kelimutu_crater_lakes bed :: chapter CH_02 :: care_tier standard :: emotional_register grief"
}
],
"target_scalars": {
"total_length_s": 88.346,
"section_count": 8,
"global_bpm": 143.55,
"baseline_bpm": 143.55,
"tempo_dispersion_bpm": 51.673,
"onset_type": "fade_in",
"time_to_hook_s": 0.139,
"time_to_full_texture_s": 22.105,
"flow_peak_position": 0.7179,
"mean_lanes": 3.686,
"max_lanes": 9,
"silence_fraction": 0.0618,
"drops_per_min": 20.3744,
"raises_per_min": 14.9413,
"breaks_per_min": 1.3583,
"tuttis_per_min": 1.3583,
"complexity_steps_per_min": 19.0161,
"boundary_step_share": 0.321,
"dynamic_range_db": 35.84,
"dynamics_peak_position": 0.0,
"note_rate_per_s": 0.894,
"melodic_salience": 0.8825,
"spectral_centroid_hz": 629.7,
"key": "E",
"mode": "major",
"signature_colour": "dark"
},
"caption_trace": [
{
"clause": "An original instrumental cue for a volcanic crater-lake rim at altitude, cold air off the water, a lone depleted child moving through it with nothing.",
"dimension": "scene (canon)",
"target_value": "docs/spine/CH_02.md:234",
"reason": "the chapter's own music_mood row is what the cue must serve; the target card supplies the grammar, never the subject",
"priority": 0
},
{
"clause": "About 88 seconds long, in 8 clearly different sections.",
"dimension": "duration_and_structure.total_length_s / section_count",
"target_value": [
88.3,
8
],
"reason": "form is the first thing a class specification fixes",
"priority": 2
},
{
"clause": "Very sparse: typically only two or three things sound at once, with the upper register left empty for long stretches.",
"dimension": "lane_analysis.simultaneity_curve.mean / max",
"target_value": [
3.69,
9
],
"reason": "section 11.1: the simultaneity curve IS the arrangement's story",
"priority": 1
},
{
"clause": "The full texture must not arrive until about 22 seconds in.",
"dimension": "opening_and_flow_law.time_to_full_texture_s",
"target_value": 22.1,
"reason": "section 11.4's time-to-full-texture -- the single lever that separates a cue that opens its hand from one that holds material back",
"priority": 1
},
{
"clause": "Stop the music completely several times. Roughly 6 percent of the running time is true silence or near silence, held long enough to be heard as a gap and not as a dip.",
"dimension": "silence_fraction",
"target_value": 0.0618,
"reason": "measured a 0.017 silence budget against a target of 0.062 -- too continuous",
"revision": true,
"priority": 1,
"replaces": "Rest is a real part of it: roughly 6 percent of the running time is near-silence, in held gaps rather than a fade."
},
{
"clause": "Instrumentation is generic orchestral and untuned percussion only: no gamelan, no gong ensemble, no bonang, no named regional instrument of any living tradition.",
"dimension": "care doctrine (CVD 17.1)",
"target_value": "build/audio/attended_care_release.json :: what_reframed_means",
"reason": "the attended care read's standing reframe: generate the PLACE, never a living culture's instrument, with the denial stated positively in the prompt",
"priority": 0
},
{
"clause": "Instrumental only, no vocals, no singing, no spoken word.",
"dimension": "generation invariant",
"target_value": "instrumental=true",
"reason": "the lane produces beds and draft realisations, never vocal takes",
"priority": 0
}
],
"caption_chars": 772,
"caption_char_budget": 774,
"caption_estimated_tokens": 188,
"caption_token_budget": 189,
"clauses_shed_to_fit_encoder_window": [
{
"dimension": "spectral_centroid_hz",
"priority": 6,
"clause": "Darker: weight the low and low-mid register, keep the bright top end out of it."
},
{
"dimension": "note_rate_per_s",
"priority": 5,
"clause": "Long notes: the material moves slowly, with sustained tones rather than busy figuration."
},
{
"dimension": "opening_and_flow_law.onset_type",
"priority": 5,
"clause": "It fades up from nothing rather than starting on a hit."
},
{
"dimension": "dynamic_range_db",
"priority": 4,
"clause": "Wide dynamics: the quiet passages sit at least 36 dB below the loudest, and nothing is compressed flat."
},
{
"dimension": "breaks_per_min",
"priority": 3,
"clause": "Cut to near-silence 2 times, each time letting the sound fall away almost completely before a single element brings it back."
},
{
"dimension": "tuttis_per_min",
"priority": 3,
"clause": "Bring the whole ensemble in together about 2 times, every part playing at once, each arrival held long enough to register."
},
{
"dimension": "tempo_and_time_grammar.global_bpm / tempo_function.dispersion_bpm",
"priority": 3,
"clause": "Around 144 beats per minute with a pulse that flexes rather than sitting on a click."
},
{
"dimension": "complexity_steps_per_min",
"priority": 2,
"clause": "Keep changing what is present: add or take away an instrument, a rhythmic subdivision or a harmony every few bars, so the texture is never the same for long."
},
{
"dimension": "dynamics_peak_position",
"priority": 2,
"clause": "The loudest passage is around 0 percent through and the ending is quieter than the middle -- it does not finish at its maximum."
}
],
"sampler_derivation": {
"duration": "= target total_length_s, rounded",
"bpm": "= target global_bpm, rounded",
"keyscale": "= target key + mode",
"seed": "the roster seed, HELD CONSTANT across revisions, so a revision is a caption change and not a reroll -- the improvement has to come from the spec",
"revision_overrides": {}
}
},
{
"sample_id": "MPX_CH02_BATTLE",
"revision": 2,
"purpose_class": "BATTLE",
"target_card": "UNMATCHED__08_-_Airfield_action",
"target_title": "Airfield action -- Deus Ex, Alexander Brandon",
"target_why": "A combat cue with every dimension non-degenerate: 81 s, mean 6.96 lanes, silence down to 1.1%, 8 tutti arrivals, 9 drops and 6 raises, and a complexity step map with a real boundary share. The denser action cue in the same corpus (12 - Hong Kong action) measures 0 complexity steps, which makes it a weaker target: a dimension that reads zero cannot show a distance shrinking.",
"canon": [
{
"file": "docs/spine/CH_02.md",
"line": 234,
"text": "music_mood -- ... combat-cue register for the stalk"
},
{
"file": "docs/spine/CH_02.md",
"line": 241,
"text": "main_boss_win_condition HP, with the fiction of the polo's hold breaking and Rangku Wera freed"
},
{
"file": "build/audio/music_cue_table.csv",
"line": 0,
"text": "MC_0035 be_0002_p1 encounter layer :: chapter CH_02 :: care_tier standard"
}
],
"target_scalars": {
"total_length_s": 80.823,
"section_count": 8,
"global_bpm": 103.36,
"baseline_bpm": 152.0,
"tempo_dispersion_bpm": 42.173,
"onset_type": "fade_in",
"time_to_hook_s": 0.116,
"time_to_full_texture_s": 12.376,
"flow_peak_position": 0.456,
"mean_lanes": 6.963,
"max_lanes": 9,
"silence_fraction": 0.0106,
"drops_per_min": 6.6813,
"raises_per_min": 4.4542,
"breaks_per_min": 0.7424,
"tuttis_per_min": 5.9389,
"complexity_steps_per_min": 5.1965,
"boundary_step_share": 0.143,
"dynamic_range_db": 6.71,
"dynamics_peak_position": 0.0,
"note_rate_per_s": 2.499,
"melodic_salience": 0.9898,
"spectral_centroid_hz": 1059.7,
"key": "C",
"mode": "major",
"signature_colour": "dark"
},
"caption_trace": [
{
"clause": "An original instrumental cue for the stalk and the fight at the crater-lake waterfall -- a hollowed hunter driven by something that empties him, closing in through volcanic jungle.",
"dimension": "scene (canon)",
"target_value": "docs/spine/CH_02.md:234",
"reason": "the chapter's own music_mood row is what the cue must serve; the target card supplies the grammar, never the subject",
"priority": 0
},
{
"clause": "Change materially about 8 times: each section brings in or removes an element so the seam is audible.",
"dimension": "section_count",
"target_value": 8,
"reason": "measured 6 sections against a target of 8 -- too static",
"revision": true,
"priority": 2,
"replaces": "About 81 seconds long, in 8 clearly different sections."
},
{
"clause": "Fill the arrangement out: keep six or more layers sounding together for most of the running time.",
"dimension": "mean_lanes",
"target_value": 6.963,
"reason": "measured 2.88 simultaneous lanes against a target of 6.963 -- too thin",
"revision": true,
"priority": 1,
"replaces": "Full and continuous: seven or more layers sound together through most of it."
},
{
"clause": "The full texture must not arrive until about 12 seconds in.",
"dimension": "opening_and_flow_law.time_to_full_texture_s",
"target_value": 12.4,
"reason": "section 11.4's time-to-full-texture -- the single lever that separates a cue that opens its hand from one that holds material back",
"priority": 1
},
{
"clause": "Keep it continuous; do not leave long empty gaps.",
"dimension": "silence_fraction",
"target_value": 0.0106,
"reason": "measured a 0.041 silence budget against a target of 0.011 -- too gappy",
"revision": true,
"priority": 1,
"replaces": "It rarely stops; near-silence is almost absent."
},
{
"clause": "Instrumentation is generic orchestral and untuned percussion only: no gamelan, no gong ensemble, no named regional instrument of any living tradition.",
"dimension": "care doctrine (CVD 17.1)",
"target_value": "build/audio/attended_care_release.json :: what_reframed_means",
"reason": "the attended care read's standing reframe: generate the PLACE, never a living culture's instrument, with the denial stated positively in the prompt",
"priority": 0
},
{
"clause": "Instrumental only, no vocals, no singing, no spoken word.",
"dimension": "generation invariant",
"target_value": "instrumental=true",
"reason": "the lane produces beds and draft realisations, never vocal takes",
"priority": 0
}
],
"caption_chars": 699,
"caption_char_budget": 774,
"caption_estimated_tokens": 170,
"caption_token_budget": 189,
"clauses_shed_to_fit_encoder_window": [
{
"dimension": "register.octave_band_profile / timbre.signature_colour",
"priority": 6,
"clause": "Weighted toward bass, low bass, sub-bass; a dark colour overall."
},
{
"dimension": "opening_and_flow_law.onset_type",
"priority": 5,
"clause": "It fades up from nothing rather than starting on a hit."
},
{
"dimension": "texture_and_density.melodic_salience",
"priority": 5,
"clause": "A clear foreground melody carries it."
},
{
"dimension": "dynamic_range_db",
"priority": 4,
"clause": "Level and driving: it stays at one intensity once it is moving, without dropping away to quiet passages."
},
{
"dimension": "tuttis_per_min",
"priority": 3,
"clause": "Bring the whole ensemble in together about 8 times, every part playing at once, each arrival held long enough to register."
},
{
"dimension": "tempo_and_time_grammar.global_bpm / tempo_function.dispersion_bpm",
"priority": 3,
"clause": "Around 103 beats per minute with a pulse that flexes rather than sitting on a click."
},
{
"dimension": "drops_per_min",
"priority": 3,
"clause": "Fewer sudden drops: about 9 in the whole cue, each one placed and meant."
},
{
"dimension": "raises_per_min",
"priority": 3,
"clause": "Fewer builds: about 6 in the whole cue, each one placed and meant."
},
{
"dimension": "complexity_steps_per_min",
"priority": 2,
"clause": "Keep changing what is present: add or take away an instrument, a rhythmic subdivision or a harmony every few bars, so the texture is never the same for long."
},
{
"dimension": "dynamics_peak_position",
"priority": 2,
"clause": "The loudest passage is around 0 percent through and the ending is quieter than the middle -- it does not finish at its maximum."
}
],
"sampler_derivation": {
"duration": "= target total_length_s, rounded",
"bpm": "= target global_bpm, rounded",
"keyscale": "= target key + mode",
"seed": "the roster seed, HELD CONSTANT across revisions, so a revision is a caption change and not a reroll -- the improvement has to come from the spec",
"revision_overrides": {}
}
},
{
"sample_id": "MPX_TITLE_MAIN",
"revision": 2,
"purpose_class": "CREDITS_TRIUMPH",
"target_card": "EX_110",
"target_title": "Want You Gone -- Portal 2, Jonathan Coulton",
"target_why": "The one ACQUIRED-and-MATCHED card in the corpus (CORPUS.json row EX_110, state ACQUIRED), so its purpose class is the corpus's own and not this lane's guess. Targeted for FORM and EVENT GRAMMAR only -- 7 sections, 29 tutti arrivals, 26 drops, 6 breaks, hook at 0.04 s, energy peak at 38% -- never for its vocal, which an instrumental generation cannot and must not reproduce.",
"canon": [
{
"file": "docs/proposals/music/MASTERPIECE_PROGRAM.md",
"line": 92,
"text": "Credits / triumph -- the summation form, motif recall economy, vocal deployment"
}
],
"target_scalars": {
"total_length_s": 141.584,
"section_count": 7,
"global_bpm": 99.38,
"baseline_bpm": 198.77,
"tempo_dispersion_bpm": 48.954,
"onset_type": "fade_in",
"time_to_hook_s": 0.035,
"time_to_full_texture_s": 3.808,
"flow_peak_position": 0.3799,
"mean_lanes": 7.269,
"max_lanes": 9,
"silence_fraction": 0.0307,
"drops_per_min": 11.0182,
"raises_per_min": 6.7804,
"breaks_per_min": 2.5427,
"tuttis_per_min": 12.2895,
"complexity_steps_per_min": 9.3231,
"boundary_step_share": 0.182,
"dynamic_range_db": 10.59,
"dynamics_peak_position": 0.4829,
"note_rate_per_s": 2.232,
"melodic_salience": 0.8372,
"spectral_centroid_hz": 2351.9,
"key": "A",
"mode": "major",
"signature_colour": "warm"
},
"caption_trace": [
{
"clause": "An original instrumental cue for a summation over the end credits: the arc's material recalled and answered, warm and forward-moving, an ending that earns its brightness.",
"dimension": "scene (canon)",
"target_value": "docs/proposals/music/MASTERPIECE_PROGRAM.md:92",
"reason": "the chapter's own music_mood row is what the cue must serve; the target card supplies the grammar, never the subject",
"priority": 0
},
{
"clause": "Change materially about 7 times: each section brings in or removes an element so the seam is audible.",
"dimension": "section_count",
"target_value": 7,
"reason": "measured 5 sections against a target of 7 -- too static",
"revision": true,
"priority": 2,
"replaces": "About 142 seconds long, in 7 clearly different sections."
},
{
"clause": "Fill the arrangement out: keep six or more layers sounding together for most of the running time.",
"dimension": "mean_lanes",
"target_value": 7.269,
"reason": "measured 4.712 simultaneous lanes against a target of 7.269 -- too thin",
"revision": true,
"priority": 1,
"replaces": "Full and continuous: seven or more layers sound together through most of it."
},
{
"clause": "The full texture must not arrive until about 4 seconds in.",
"dimension": "opening_and_flow_law.time_to_full_texture_s",
"target_value": 3.8,
"reason": "section 11.4's time-to-full-texture -- the single lever that separates a cue that opens its hand from one that holds material back",
"priority": 1
},
{
"clause": "Energy is held back and peaks about 38 percent of the way through, not at the start.",
"dimension": "opening_and_flow_law.flow_curve.peak_position_normalised",
"target_value": 0.38,
"reason": "the flow curve's archetypal shape per class (section 11.4)",
"priority": 2
},
{
"clause": "A few short near-silent gaps, about 3 percent of the running time.",
"dimension": "event_grammar.silence_budget.fraction",
"target_value": 0.0307,
"reason": "section 11.2 treats rest as a FIRST-CLASS parameter",
"priority": 1
},
{
"clause": "Instrumental only, no vocals, no singing, no spoken word.",
"dimension": "generation invariant",
"target_value": "instrumental=true",
"reason": "the lane produces beds and draft realisations, never vocal takes",
"priority": 0
}
],
"caption_chars": 639,
"caption_char_budget": 774,
"caption_estimated_tokens": 156,
"caption_token_budget": 189,
"clauses_shed_to_fit_encoder_window": [
{
"dimension": "spectral_centroid_hz",
"priority": 6,
"clause": "Brighter: let the upper register and presence band carry."
},
{
"dimension": "opening_and_flow_law.onset_type",
"priority": 5,
"clause": "It fades up from nothing rather than starting on a hit."
},
{
"dimension": "texture_and_density.melodic_salience",
"priority": 5,
"clause": "A clear foreground melody carries it."
},
{
"dimension": "dynamic_range_db",
"priority": 4,
"clause": "Level and driving: it stays at one intensity once it is moving, without dropping away to quiet passages."
},
{
"dimension": "breaks_per_min",
"priority": 3,
"clause": "Cut to near-silence 6 times, each time letting the sound fall away almost completely before a single element brings it back."
},
{
"dimension": "tuttis_per_min",
"priority": 3,
"clause": "Bring the whole ensemble in together about 29 times, every part playing at once, each arrival held long enough to register."
},
{
"dimension": "tempo_and_time_grammar.global_bpm / tempo_function.dispersion_bpm",
"priority": 3,
"clause": "Around 99 beats per minute with a pulse that flexes rather than sitting on a click."
},
{
"dimension": "drops_per_min",
"priority": 3,
"clause": "Fewer sudden drops: about 26 in the whole cue, each one placed and meant."
},
{
"dimension": "raises_per_min",
"priority": 3,
"clause": "Fewer builds: about 16 in the whole cue, each one placed and meant."
},
{
"dimension": "complexity_steps_per_min",
"priority": 2,
"clause": "Keep changing what is present: add or take away an instrument, a rhythmic subdivision or a harmony every few bars, so the texture is never the same for long."
}
],
"sampler_derivation": {
"duration": "= target total_length_s, rounded",
"bpm": "= target global_bpm, rounded",
"keyscale": "= target key + mode",
"seed": "the roster seed, HELD CONSTANT across revisions, so a revision is a caption change and not a reroll -- the improvement has to come from the spec",
"revision_overrides": {}
}
}
],
"jobs": [
{
"artifact_id": "MPX_CH02_EXPLORE_r2",
"chapter_id": "CH_02",
"care_tier": "standard",
"caption": "An original instrumental cue for a volcanic crater-lake rim at altitude, cold air off the water, a lone depleted child moving through it with nothing. About 88 seconds long, in 8 clearly different sections. Very sparse: typically only two or three things sound at once, with the upper register left empty for long stretches. The full texture must not arrive until about 22 seconds in. Stop the music completely several times. Roughly 6 percent of the running time is true silence or near silence, held long enough to be heard as a gap and not as a dip. Instrumentation is generic orchestral and untuned percussion only: no gamelan, no gong ensemble, no bonang, no named regional instrument of any living tradition. Instrumental only, no vocals, no singing, no spoken word.",
"duration": 88,
"bpm": 144,
"keyscale": "E major",
"timesignature": "4/4",
"instrumental": true,
"inference_steps": 60,
"guidance_scale": 7.5,
"shift": 3.0,
"seed": 7101,
"generation_tier": "region_standard",
"render_tier": "REVIEW",
"audio_format": "wav"
},
{
"artifact_id": "MPX_CH02_BATTLE_r2",
"chapter_id": "CH_02",
"care_tier": "standard",
"caption": "An original instrumental cue for the stalk and the fight at the crater-lake waterfall -- a hollowed hunter driven by something that empties him, closing in through volcanic jungle. Change materially about 8 times: each section brings in or removes an element so the seam is audible. Fill the arrangement out: keep six or more layers sounding together for most of the running time. The full texture must not arrive until about 12 seconds in. Keep it continuous; do not leave long empty gaps. Instrumentation is generic orchestral and untuned percussion only: no gamelan, no gong ensemble, no named regional instrument of any living tradition. Instrumental only, no vocals, no singing, no spoken word.",
"duration": 81,
"bpm": 103,
"keyscale": "C major",
"timesignature": "4/4",
"instrumental": true,
"inference_steps": 60,
"guidance_scale": 7.5,
"shift": 3.0,
"seed": 7201,
"generation_tier": "region_standard",
"render_tier": "REVIEW",
"audio_format": "wav"
},
{
"artifact_id": "MPX_TITLE_MAIN_r2",
"chapter_id": "",
"care_tier": "standard",
"caption": "An original instrumental cue for a summation over the end credits: the arc's material recalled and answered, warm and forward-moving, an ending that earns its brightness. Change materially about 7 times: each section brings in or removes an element so the seam is audible. Fill the arrangement out: keep six or more layers sounding together for most of the running time. The full texture must not arrive until about 4 seconds in. Energy is held back and peaks about 38 percent of the way through, not at the start. A few short near-silent gaps, about 3 percent of the running time. Instrumental only, no vocals, no singing, no spoken word.",
"duration": 142,
"bpm": 99,
"keyscale": "A major",
"timesignature": "4/4",
"instrumental": true,
"inference_steps": 60,
"guidance_scale": 7.5,
"shift": 3.0,
"seed": 7301,
"generation_tier": "region_standard",
"render_tier": "REVIEW",
"audio_format": "wav"
}
],
"_bench_note": "IDENTICAL to build/audio/proof/spec_r2.json in every job field — same captions, same seeds, same bpm/keyscale/duration/steps/guidance. The ONLY change is required_generator, so the checkpoint is the single moving part in the A/B against the three samples Josh rejected."
}