{
  "protocol": "Three independent training seeds (0,1,2), fixed 1437/360 stratified split, 12000 optimizer updates each. Primary shipped checkpoint is seed 0 EMA; not chosen by gallery or cross-seed maximum.",
  "runs": [
    {
      "initial_validation_mse": 1.996389389038086,
      "zero_velocity_mse": 1.7100704908370972,
      "negative_input_mse": 1.996389389038086,
      "validation_mse": 0.4002799987792969,
      "raw_validation_mse": 0.4015161693096161,
      "ema_validation_mse": 0.4002799987792969,
      "wrong_label_mse": 0.7228743433952332,
      "null_label_mse": 0.517828643321991,
      "one_inner_step_mse": 1.7464349269866943,
      "selected_weights": "ema",
      "parameter_count": 10674,
      "train_images": 1437,
      "validation_images": 360,
      "optimizer_steps": 12000,
      "seed": 0,
      "wall_seconds": 233.75435167700005,
      "dataset": "scikit-learn load_digits / UCI Optical Recognition of Handwritten Digits; 1797 8x8 images",
      "scope": "8x8 grayscale class-conditioned proof, NOT text-to-image or high-resolution evidence",
      "environment": {
        "python": "3.13.5",
        "torch": "2.10.0+cpu",
        "device": "cpu",
        "threads": 1
      },
      "recogniser_real_validation_accuracy": 0.9722222222222222,
      "generated_label_accuracy": 0.872,
      "generated_nearest_train_mse_mean": 0.1328294724225998,
      "generated_near_exact_train_copy_fraction_mse_lt_1e_6": 0.0,
      "generation_count": 500
    },
    {
      "initial_validation_mse": 1.996389389038086,
      "zero_velocity_mse": 1.7100704908370972,
      "negative_input_mse": 1.996389389038086,
      "validation_mse": 0.3948173522949219,
      "raw_validation_mse": 0.39585936069488525,
      "ema_validation_mse": 0.3948173522949219,
      "wrong_label_mse": 0.7173572182655334,
      "null_label_mse": 0.5159406661987305,
      "one_inner_step_mse": 1.8324545621871948,
      "selected_weights": "ema",
      "parameter_count": 10674,
      "train_images": 1437,
      "validation_images": 360,
      "optimizer_steps": 12000,
      "seed": 1,
      "wall_seconds": 228.88593255500007,
      "dataset": "scikit-learn load_digits / UCI Optical Recognition of Handwritten Digits; 1797 8x8 images",
      "scope": "8x8 grayscale class-conditioned proof, NOT text-to-image or high-resolution evidence",
      "environment": {
        "python": "3.13.5",
        "torch": "2.10.0+cpu",
        "device": "cpu",
        "threads": 1
      },
      "recogniser_real_validation_accuracy": 0.9722222222222222,
      "generated_label_accuracy": 0.874,
      "generated_nearest_train_mse_mean": 0.13002696633338928,
      "generated_near_exact_train_copy_fraction_mse_lt_1e_6": 0.0,
      "generation_count": 500
    },
    {
      "initial_validation_mse": 1.996389389038086,
      "zero_velocity_mse": 1.7100704908370972,
      "negative_input_mse": 1.996389389038086,
      "validation_mse": 0.3970016539096832,
      "raw_validation_mse": 0.3995174467563629,
      "ema_validation_mse": 0.3970016539096832,
      "wrong_label_mse": 0.7188294529914856,
      "null_label_mse": 0.519249677658081,
      "one_inner_step_mse": 1.8308186531066895,
      "selected_weights": "ema",
      "parameter_count": 10674,
      "train_images": 1437,
      "validation_images": 360,
      "optimizer_steps": 12000,
      "seed": 2,
      "wall_seconds": 226.02392982900005,
      "dataset": "scikit-learn load_digits / UCI Optical Recognition of Handwritten Digits; 1797 8x8 images",
      "scope": "8x8 grayscale class-conditioned proof, NOT text-to-image or high-resolution evidence",
      "environment": {
        "python": "3.13.5",
        "torch": "2.10.0+cpu",
        "device": "cpu",
        "threads": 1
      },
      "recogniser_real_validation_accuracy": 0.9722222222222222,
      "generated_label_accuracy": 0.86,
      "generated_nearest_train_mse_mean": 0.13300356268882751,
      "generated_near_exact_train_copy_fraction_mse_lt_1e_6": 0.0,
      "generation_count": 500
    }
  ],
  "generated_label_accuracy_mean": 0.8686666666666667,
  "generated_label_accuracy_sample_std": 0.007571877794400371,
  "generated_label_accuracy_min": 0.86,
  "generated_label_accuracy_max": 0.874,
  "limitations": [
    "8x8 grayscale digits only; not caption-conditioned",
    "500 generated samples per seed, same evaluation noise bank; recognizer is logistic regression trained on real training images",
    "Real-image recognizer accuracy is 97.22%, not an infallible perceptual oracle",
    "Validation split used to choose raw versus EMA, not an untouched final test set",
    "One-inner-step ablation is a test-time intervention, not a separately trained matched-compute baseline",
    "Near-exact pixel-copy check does not establish absence of memorization",
    "No physical GPU available; CPU training times are functional evidence, not GPU throughput benchmarks"
  ]
}