{
  "schema_version": "demo_quality_contract.v1",
  "model_card_promise": "Text-to-music generation from a prompt with optional lyrics, producing stereophonic 48 kHz audio. Turbo-distilled 5B DiT with 8-step flow-matching inference.",
  "demo_task": "text-to-audio",
  "primary_user_flow": "User visits Generate tab, types a music description prompt, optionally adds lyrics, sets duration/seed, clicks Generate, and receives a playable WAV output.",
  "examples_provided": true,
  "examples": [
    { "prompt": "An upbeat synthwave track with driving drums and a catchy lead", "lyrics": "", "audio_duration": 30, "seed": 42 },
    { "prompt": "A soft piano ballad with gentle strings", "lyrics": "Stars above us\nQuiet nights\nForever here", "audio_duration": 20, "seed": 7 },
    { "prompt": "Energetic drum and bass with deep basslines", "lyrics": "", "audio_duration": 15, "seed": 99 }
  ],
  "canonical_smoke_example": {
    "api_name": "/generate",
    "inputs": { "prompt": "An upbeat synthwave track with driving drums and a catchy lead", "lyrics": "", "audio_duration": 10, "seed": 42 },
    "expected_output_type": "audio/wav",
    "reason": "10-second audio at 8 steps is cheap, exercises full pipeline, and can be validated by Gradio client returning a WAV file path."
  },
  "real_inference_required": true,
  "real_inference_implemented": true,
  "fallback_or_diagnostic_only": false,
  "limitations_disclosed": [
    "Cold-start download and weight loading may delay first generation.",
    "Guidance scale > 1.0 is ignored (turbo-distilled checkpoint).",
    "VAE tiling is enabled but very long durations may still risk OOM; reducing duration keeps the app stable."
  ],
  "promise_fulfillment_risk": "low"
}