{
  "strict_load": true,
  "source_tensor_bytes_equal": true,
  "checks": [
    {
      "name": "649_source_tensor_payloads",
      "passed": true
    },
    {
      "name": "native_strict_load",
      "passed": true
    },
    {
      "name": "official_scale_schedule",
      "passed": true
    },
    {
      "name": "multiscale_token_geometry",
      "passed": true,
      "shape": [
        1,
        817
      ]
    },
    {
      "name": "ID_bounds",
      "passed": true
    },
    {
      "name": "residual_lookup_bound",
      "passed": true,
      "max_abs": 4.76837158203125e-07
    },
    {
      "name": "semantic_output",
      "passed": true,
      "shape": [
        1,
        256,
        512
      ]
    },
    {
      "name": "RGB_output",
      "passed": true,
      "shape": [
        1,
        3,
        256,
        256
      ],
      "mse": 0.00044855583109892905
    },
    {
      "name": "semantic_decode_code_lookup_exact",
      "passed": true
    },
    {
      "name": "RGB_decode_code_lookup_exact",
      "passed": true
    }
  ],
  "smoke": {
    "passed": true,
    "device": "cpu",
    "dtype": "float32"
  },
  "elapsed_seconds": 30.36184666599729,
  "limitations": [
    "Synthetic 256px native codec fixtures, not image quality benchmarks or complete VQA/T2I pipelines.",
    "Includes the trained CLIP visual branch, semantic and pixel codebooks, semantic decoder and enhanced RGB decoder. No CLIP text model or external pretrained download required.",
    "Native straight-through arithmetic can differ slightly from direct lookup; canonical lookup and decode-code paths are compared exactly."
  ]
}