{
  "strict_load": true,
  "source_tensor_bytes_equal": true,
  "checks": [
    {
      "name": "684_source_tensor_payloads",
      "passed": true
    },
    {
      "name": "native_strict_load",
      "passed": true
    },
    {
      "name": "original_VAE_scale",
      "passed": true
    },
    {
      "name": "64x64_native_encode_exact",
      "passed": true,
      "shape": [
        1,
        4,
        4,
        16
      ]
    },
    {
      "name": "64x64_native_decode_exact",
      "passed": true
    },
    {
      "name": "64x64_finite_reconstruction",
      "passed": true,
      "mse": 0.0004284594615455717
    },
    {
      "name": "64x64_masked_encoder_geometry",
      "passed": true,
      "shape": [
        1,
        72,
        768
      ]
    },
    {
      "name": "64x64_masked_decoder_geometry",
      "passed": true
    },
    {
      "name": "64x64_language_projections",
      "passed": true
    },
    {
      "name": "64x96_native_encode_exact",
      "passed": true,
      "shape": [
        1,
        4,
        6,
        16
      ]
    },
    {
      "name": "64x96_native_decode_exact",
      "passed": true
    },
    {
      "name": "64x96_finite_reconstruction",
      "passed": true,
      "mse": 0.0003114198916591704
    },
    {
      "name": "64x96_masked_encoder_geometry",
      "passed": true,
      "shape": [
        1,
        76,
        768
      ]
    },
    {
      "name": "64x96_masked_decoder_geometry",
      "passed": true
    },
    {
      "name": "64x96_language_projections",
      "passed": true
    },
    {
      "name": "diffusion_prediction_and_variance",
      "passed": true
    },
    {
      "name": "native_100step_sampling_cpu",
      "passed": true
    },
    {
      "name": "native_CFG_sampling_cpu",
      "passed": true
    }
  ],
  "smoke": {
    "passed": true,
    "device": "cpu",
    "dtype": "float32"
  },
  "elapsed_seconds": 7.45948820799822,
  "limitations": [
    "Synthetic 64px square/rectangular VAE/MAR fixtures and two-token native diffusion sampling; not image quality or downstream VQA benchmarks.",
    "The trained Qwen model is excluded; complete language-conditioned generation still requires its hidden states and the official masked generation schedule.",
    "Sampling tensor allocations follow the input device instead of hardcoded CUDA; native probability distributions and sampler equations are retained."
  ]
}