ElevenLabs Music v2 (Josh's Starter tier) - prompt CUE-MC0035_LiangBuaTension_B - EXPLORATION - 120.0 s - judged 2026-08-11 by benchmark-judge/1.0.0
Status: UNGRADED. Nobody has heard this but the operator who downloaded it.
What the wall read
Structural yield UNAVAILABLE: only 3 novelty boundaries on the card; the ratio is not meaningful below four
Melody bands (audio tier): floor FAILS against EXPLORATION (n=8).
BELOW the corpus p10 on: note_density_per_s, note_density_per_sounding_s, interval_class_count, contour_turning_rate, contour_depth, phrase_peak_position, range_semitones. This is the 'a few notes slowly played back to back' failure, measured.
Above the corpus p90 on: step_share, distinct_durations - not a failure by itself; on the interval axes it is the measurable form of 'they clash'.
Loudness (INFORMATIONAL, not a gate): -14.66 LUFS against a corpus EXPLORATION median of -12.18; range 16.43 LU; true peak 0.01 dBTP.
Quotation screen: CLEAR over 46 extracted intervals against 13 corpus hook signatures.
2 of 7 instruments that could decide say pass.
What this did NOT measure
harmony_fit (NEEDS-SCORE): HARMONY FIT IS A SCORE INSTRUMENT AND HAS NO AUDIO TIER. Chord-tone agreement, dissonance classification and parallel-perfect detection all require knowing WHICH NOTES ARE SOUNDING, and docs/proposals/music/craft_research/COUNTERPOINT_AND_VOICE_LEADING.md:559-569 puts exactly those on the not-measurable-from-audio list (GAP 1, polyphonic transcription). Pass --plan. A card-only or audio-only tier here would be an invented proxy for the one thing the research says cannot be recovered from a render.
Recoverable for this engine: False. audio only; the NEEDS-SCORE instruments are unrecoverable for this engine and stay blind spots on every row
(none exists) (UNMEASURED): carrier migration -- the melody passed BETWEEN instruments on its returns -- no instrument in the battery identifies WHICH instrument is carrying the tune at a given moment. instr_melody_bands extracts ONE predominant line and folds octave doublings; instr_hook_presence sees the motif return but not who plays it. Source separation would be the honest route and nothing in this program uses it on the melodic layer.
Recoverable for this engine: False. the single clause Josh named most specifically on 2026-08-09 is graded by EAR ONLY. No wall verdict may be read as evidence for or against it.
(none exists) (UNMEASURED): does the ending SETTLE and hand off to the next track -- feature_rig.loop_seam reads the seam and whether a fade was baked in, which is a different question from whether the last phrase resolves. Nothing reads cadence.
Recoverable for this engine: False. graded by ear. The seam measurement is reported and must not stand in for it.
loudness_informational is excluded from the verdict: it fails the corpus; see measurements.loudness_informational.why_not_a_gate
The honest label, forwarded verbatim from the battery
Nothing here is a judgement of quality and no listening test stands behind any of it. Three of the four instruments looked for Josh's defect in their own domain and did not find it: the cycle law passes 17 of the 18 tracks he rejected (within-album AUC 0.501), the noise detector fires on 0 of 18 across both tiers with nothing clearing the multiplicity bar, and the layer census saturates near 8 streams where a real cue runs 24 to 40 parts, so it cannot adjudicate the layer clause at all. Sensory roughness came back at chance (0.4963). Structural yield is the axis that separates, and it separates our generator from real music, not good music from filler.
Licence posture
PAID-COMMERCIAL, BUT STUDIO GAMES ARE CARVED OUT BELOW ENTERPRISE. Shippable today: False.
Evidence: docs/licence_records/elevenlabs/READ_RECORD_2026-08-09.md -- the Model-Specific Terms (Last Updated 26 May 2026) define Studio Games as commercialised video games on more than one platform, and elevenlabs.io/music-api states 'Film, TV, and large studio game rights require an Enterprise plan'. Starter is JUDGING-ONLY for this project
Outstanding read: the Music Commercial Rights table itself is client-rendered and did not fetch; read it in the browser and pin the Starter row verbatim
Generated by harness/site/structure_site.py — the URL path is the repo path. review root