{
  "status": "passed_negative_candidate",
  "generated_at": "2026-09-10T02:17:38.099117+00:00",
  "results": {
    "original": {
      "exact_to_saved_output": true,
      "samples": 900,
      "independent_launches": 1,
      "repeat_mean_ms": [
        0.4883512532711029,
        0.48453098654747007,
        0.48449557373921076
      ],
      "mean_of_repeat_means_ms": 0.48579260451926126
    },
    "channels_last_inclusive": {
      "exact_to_saved_output": true,
      "samples": 900,
      "independent_launches": 1,
      "repeat_mean_ms": [
        0.595068374077479,
        0.5947258665164312,
        0.5943578654527664
      ],
      "mean_of_repeat_means_ms": 0.5947173686822256
    }
  },
  "candidate_time_ratio": 1.2242207130154974,
  "input_sha256": "1e47965836d719f164296b5c8d52cac8c22a10d9ff8c5f4fdfa48498909f8856",
  "allocated": "6",
  "scope": "One process, same real input and NCDHW output, three interleaved 300-sample batches per mode. All packing/conversion costs included. Local eager CUDA-event timing, no profiler. Samples correlated, not three independent launches or a model P99/stability study. No integrated speedup accepted; a negative layout candidate is not evidence that all VAE layout optimization is impossible.",
  "sha256": {
    "script.py": "9cdc2bf7b264c56fc1d9103ae2a2b665bb454d5b1740a22b8b3ebb710da8d864",
    "original_output.pt": "7c49d894327fef3454ae0edeaeb90653eb4ea04e2767dec2d5320f9e60fb79f1",
    "contract.json": "f174fc263b20c7d908ba0a18d6ae1fbec34f1d0cbef227386278945f483934a4",
    "channels_last_inclusive_output.pt": "c68edf32896e041d6e94c73c8388910410834a412519a06b617495fe0db0c69f",
    "gpu_environment.txt": "c639339ff922a5cdfaee97260f170e5e54797c41884fae395164a515dca1fec1",
    "manifest.json": "89025d6fcbbd38a6bef0f1a64941a5d18a87868e4651a0ace1aae8b4b0d9f7be"
  }
}