{
  "schema_version": "inference_contract.v1",
  "full_inference_implemented": true,
  "inference_strategy": "diffusers_pipeline",
  "health_endpoint": "/health",
  "primary_api_name": "/generate",
  "expected_output_type": "audio",
  "validation_level": "live_gradio_client_smoke",
  "requires_gpu": true,
  "recommended_target_space_hardware": "zero-a10g",
  "estimated_vram": "12-18 GB for 30s generation with bf16 and vae tiling",
  "blockers_count": 0,
  "model_family": "diffusers_full_pipeline",
  "artifact_role": "full_model",
  "loader_strategy": "DiffusionPipeline_or_family_specific_from_pretrained",
  "app_runtime_contract_version": "app_runtime_contract.v198_26_30",
  "health_required_fields": ["status", "model_ready", "pipeline_ready", "model_family", "loader_strategy", "last_error", "expected_output_type"],
  "generate_not_ready_behavior": "raises RuntimeError with concrete last_error or generation-specific exception",
  "recipe_conformance_notes": [
    "Loader uses AceStepPipeline.from_pretrained per model_card/recipe.",
    "pipe is loaded at module level with .to('cuda') and vae tiling enabled.",
    "Inference is wrapped with @spaces.GPU(duration=120).",
    "Health is cheap, does not load model, and returns all required contract fields.",
    "Exceptions from model load and generation are surfaced in last_error and RuntimeError respectively."
  ]
}