{
  "evidence_status": "measured fixed-subset teaching experiment; not a benchmark",
  "sealed_test_contract": {
    "regime_and_epoch_selected_from": "validation only",
    "selection_order": [
      "maximum validation accuracy",
      "fewest trainable parameters",
      "minimum validation loss",
      "earliest epoch"
    ],
    "selected_regime": "all",
    "selected_epoch": 13,
    "test_evaluations": 1,
    "test_evaluated_after_selection": true
  },
  "dataset": {
    "name": "Oxford-IIIT Pet",
    "dataset_page": "https://www.robots.ox.ac.uk/~vgg/data/pets/",
    "images_archive": "https://thor.robots.ox.ac.uk/pets/images.tar.gz",
    "annotations_archive": "https://thor.robots.ox.ac.uk/pets/annotations.tar.gz",
    "license": "CC BY-SA 4.0; copyright remains with original image owners",
    "breeds": [
      "Abyssinian",
      "Bengal",
      "Birman",
      "Bombay",
      "British_Shorthair",
      "Egyptian_Mau"
    ],
    "official_boundary": "train and validation are disjoint subselects of official trainval; sealed test is a subselect of official test",
    "selection_population": "official split rows whose JPEG was present under --images-dir when selection-manifest.csv was bootstrapped",
    "selection_rank": "SHA256('20260812:<official_split>:<image_id>')",
    "counts": {
      "train": 72,
      "val": 36,
      "test": 36
    },
    "availability_at_bootstrap": {
      "Abyssinian": {
        "trainval_present": 30,
        "trainval_available": 30,
        "trainval_rejected_invalid": 0,
        "test_present": 27,
        "test_available": 27,
        "test_rejected_invalid": 0
      },
      "Bengal": {
        "trainval_present": 21,
        "trainval_available": 21,
        "trainval_rejected_invalid": 0,
        "test_present": 20,
        "test_available": 20,
        "test_rejected_invalid": 0
      },
      "Birman": {
        "trainval_present": 23,
        "trainval_available": 22,
        "trainval_rejected_invalid": 1,
        "test_present": 22,
        "test_available": 22,
        "test_rejected_invalid": 0
      },
      "Bombay": {
        "trainval_present": 24,
        "trainval_available": 24,
        "trainval_rejected_invalid": 0,
        "test_present": 24,
        "test_available": 24,
        "test_rejected_invalid": 0
      },
      "British_Shorthair": {
        "trainval_present": 28,
        "trainval_available": 28,
        "trainval_rejected_invalid": 0,
        "test_present": 38,
        "test_available": 38,
        "test_rejected_invalid": 0
      },
      "Egyptian_Mau": {
        "trainval_present": 26,
        "trainval_available": 26,
        "trainval_rejected_invalid": 0,
        "test_present": 23,
        "test_available": 23,
        "test_rejected_invalid": 0
      }
    },
    "official_split_hashes": {
      "trainval.txt": "408f3f609481b939c94634169e6413414b733a3faeba440cbdcc5c02142eebdc",
      "test.txt": "a5454003774ffe01f4f322756d3ba5495bae21cb30bb217ab285dbfa2bef245c"
    },
    "manifest_sha256": "5b5479a311cc2189522c478b4b2dbe014619f31831760ccadf419ddc41619f1e"
  },
  "model": {
    "architecture": "torchvision ResNet-18",
    "weights_enum": "ResNet18_Weights.IMAGENET1K_V1",
    "weights_url": "https://download.pytorch.org/models/resnet18-f37072fd.pth",
    "weights_sha256": "f37072fd47e89c5e827621c5baffa7500819f7896bbacec160b1a16c560e07ec",
    "feature_dimension": 512,
    "head": "Linear(512, 6)",
    "head_parameters": 3078,
    "shared_head_initialization_sha256": "09a00c752d93b078e60e02cd8123947e3845033e25477e84fbc4c60760e7805c"
  },
  "preprocessing": {
    "source": "ResNet18_Weights.IMAGENET1K_V1.transforms()",
    "resize_size": [
      256
    ],
    "crop_size": [
      224
    ],
    "interpolation": "InterpolationMode.BILINEAR",
    "antialias": true,
    "mean": [
      0.485,
      0.456,
      0.406
    ],
    "std": [
      0.229,
      0.224,
      0.225
    ],
    "augmentation": "none",
    "observed_shared_image": {
      "original_size_wh": [
        600,
        400
      ],
      "resized_size_wh": [
        384,
        256
      ],
      "crop_size_wh": [
        224,
        224
      ],
      "tensor_shape_chw": [
        3,
        224,
        224
      ],
      "channel_stats_after_normalization": [
        {
          "channel": "R",
          "min": -2.1007792949676514,
          "max": 1.9920369386672974,
          "mean": -0.5898733139038086,
          "std": 0.7244507074356079
        },
        {
          "channel": "G",
          "min": -1.9656862020492554,
          "max": 1.8508403301239014,
          "mean": -0.5304905772209167,
          "std": 0.6695410013198853
        },
        {
          "channel": "B",
          "min": -1.804444432258606,
          "max": 2.1345534324645996,
          "mean": -0.651914656162262,
          "std": 0.6381468176841736
        }
      ]
    }
  },
  "protocol": {
    "seed": 20260812,
    "epochs": 15,
    "batch_size": 12,
    "optimizer": "AdamW",
    "weight_decay": 0.0001,
    "learning_rates": {
      "head_all_regimes": 0.003,
      "late_layer4": 0.0003,
      "all_backbone": 3e-05
    },
    "batch_norm": "backbone kept in eval mode; buffers fixed; affine parameters train only in unfrozen layers",
    "same_training_order_across_regimes": true,
    "same_head_initialization_across_regimes": true,
    "epoch_selection": "max val accuracy, then min val loss, then earliest epoch"
  },
  "regimes": [
    {
      "regime": "probe",
      "trainable_parameters": 3078,
      "best": {
        "regime": "probe",
        "epoch": 14,
        "train_loss": 0.06518923367063205,
        "train_accuracy": 1.0,
        "val_loss": 0.34095456699530285,
        "val_accuracy": 0.9166666666666666
      },
      "history": [
        {
          "regime": "probe",
          "epoch": 1,
          "train_loss": 1.8142588933308919,
          "train_accuracy": 0.25,
          "val_loss": 1.3921135928895738,
          "val_accuracy": 0.4444444444444444
        },
        {
          "regime": "probe",
          "epoch": 2,
          "train_loss": 1.1881758868694305,
          "train_accuracy": 0.5277777777777778,
          "val_loss": 0.9760624832577176,
          "val_accuracy": 0.6944444444444444
        },
        {
          "regime": "probe",
          "epoch": 3,
          "train_loss": 0.7199392914772034,
          "train_accuracy": 0.8611111111111112,
          "val_loss": 0.8684951729244657,
          "val_accuracy": 0.6666666666666666
        },
        {
          "regime": "probe",
          "epoch": 4,
          "train_loss": 0.4570140639940898,
          "train_accuracy": 0.9166666666666666,
          "val_loss": 0.6317573852009244,
          "val_accuracy": 0.7777777777777778
        },
        {
          "regime": "probe",
          "epoch": 5,
          "train_loss": 0.360688457886378,
          "train_accuracy": 0.9166666666666666,
          "val_loss": 0.5255186822679307,
          "val_accuracy": 0.8333333333333334
        },
        {
          "regime": "probe",
          "epoch": 6,
          "train_loss": 0.24436519294977188,
          "train_accuracy": 0.9861111111111112,
          "val_loss": 0.5302794443236457,
          "val_accuracy": 0.8333333333333334
        },
        {
          "regime": "probe",
          "epoch": 7,
          "train_loss": 0.20543111364046732,
          "train_accuracy": 0.9861111111111112,
          "val_loss": 0.46042878097958034,
          "val_accuracy": 0.8333333333333334
        },
        {
          "regime": "probe",
          "epoch": 8,
          "train_loss": 0.15523658817013106,
          "train_accuracy": 1.0,
          "val_loss": 0.41534939408302307,
          "val_accuracy": 0.8888888888888888
        },
        {
          "regime": "probe",
          "epoch": 9,
          "train_loss": 0.13855032746990523,
          "train_accuracy": 1.0,
          "val_loss": 0.39353909426265293,
          "val_accuracy": 0.8611111111111112
        },
        {
          "regime": "probe",
          "epoch": 10,
          "train_loss": 0.10657061263918877,
          "train_accuracy": 1.0,
          "val_loss": 0.3669303059577942,
          "val_accuracy": 0.8888888888888888
        },
        {
          "regime": "probe",
          "epoch": 11,
          "train_loss": 0.09475106621781985,
          "train_accuracy": 1.0,
          "val_loss": 0.3601375487115648,
          "val_accuracy": 0.8888888888888888
        },
        {
          "regime": "probe",
          "epoch": 12,
          "train_loss": 0.0819233035047849,
          "train_accuracy": 1.0,
          "val_loss": 0.35185642706023323,
          "val_accuracy": 0.8888888888888888
        },
        {
          "regime": "probe",
          "epoch": 13,
          "train_loss": 0.07301352918148041,
          "train_accuracy": 1.0,
          "val_loss": 0.3495305859380298,
          "val_accuracy": 0.8888888888888888
        },
        {
          "regime": "probe",
          "epoch": 14,
          "train_loss": 0.06518923367063205,
          "train_accuracy": 1.0,
          "val_loss": 0.34095456699530285,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "probe",
          "epoch": 15,
          "train_loss": 0.06188371405005455,
          "train_accuracy": 1.0,
          "val_loss": 0.33068979448742336,
          "val_accuracy": 0.8888888888888888
        }
      ]
    },
    {
      "regime": "late",
      "trainable_parameters": 8396806,
      "best": {
        "regime": "late",
        "epoch": 4,
        "train_loss": 0.4968356390794118,
        "train_accuracy": 0.8472222222222222,
        "val_loss": 0.7534577747186025,
        "val_accuracy": 0.75
      },
      "history": [
        {
          "regime": "late",
          "epoch": 1,
          "train_loss": 2.042006492614746,
          "train_accuracy": 0.18055555555555555,
          "val_loss": 1.6768923335605197,
          "val_accuracy": 0.3611111111111111
        },
        {
          "regime": "late",
          "epoch": 2,
          "train_loss": 1.539662520090739,
          "train_accuracy": 0.4027777777777778,
          "val_loss": 1.2964441378911336,
          "val_accuracy": 0.4444444444444444
        },
        {
          "regime": "late",
          "epoch": 3,
          "train_loss": 0.937593807776769,
          "train_accuracy": 0.7083333333333334,
          "val_loss": 1.1298948625723522,
          "val_accuracy": 0.5833333333333334
        },
        {
          "regime": "late",
          "epoch": 4,
          "train_loss": 0.4968356390794118,
          "train_accuracy": 0.8472222222222222,
          "val_loss": 0.7534577747186025,
          "val_accuracy": 0.75
        },
        {
          "regime": "late",
          "epoch": 5,
          "train_loss": 0.1369476163915048,
          "train_accuracy": 0.9583333333333334,
          "val_loss": 1.0181353290875752,
          "val_accuracy": 0.75
        },
        {
          "regime": "late",
          "epoch": 6,
          "train_loss": 0.07266251238373418,
          "train_accuracy": 0.9861111111111112,
          "val_loss": 2.920265727572971,
          "val_accuracy": 0.5833333333333334
        },
        {
          "regime": "late",
          "epoch": 7,
          "train_loss": 0.06565285543911159,
          "train_accuracy": 1.0,
          "val_loss": 1.3920429415173001,
          "val_accuracy": 0.6944444444444444
        },
        {
          "regime": "late",
          "epoch": 8,
          "train_loss": 0.05218614395319795,
          "train_accuracy": 0.9722222222222222,
          "val_loss": 2.2226748731401234,
          "val_accuracy": 0.5555555555555556
        },
        {
          "regime": "late",
          "epoch": 9,
          "train_loss": 0.03971123830221283,
          "train_accuracy": 0.9861111111111112,
          "val_loss": 2.1752648750933883,
          "val_accuracy": 0.5277777777777778
        },
        {
          "regime": "late",
          "epoch": 10,
          "train_loss": 0.06224960220667223,
          "train_accuracy": 0.9861111111111112,
          "val_loss": 3.28106070889367,
          "val_accuracy": 0.4444444444444444
        },
        {
          "regime": "late",
          "epoch": 11,
          "train_loss": 0.3028248354191116,
          "train_accuracy": 0.9583333333333334,
          "val_loss": 0.9986932145224677,
          "val_accuracy": 0.6944444444444444
        },
        {
          "regime": "late",
          "epoch": 12,
          "train_loss": 0.10894174960170251,
          "train_accuracy": 0.9583333333333334,
          "val_loss": 1.630471978627611,
          "val_accuracy": 0.6111111111111112
        },
        {
          "regime": "late",
          "epoch": 13,
          "train_loss": 0.02114047948271036,
          "train_accuracy": 1.0,
          "val_loss": 1.9697204785431193,
          "val_accuracy": 0.6388888888888888
        },
        {
          "regime": "late",
          "epoch": 14,
          "train_loss": 0.006132793088909239,
          "train_accuracy": 1.0,
          "val_loss": 2.1162425989750773,
          "val_accuracy": 0.6388888888888888
        },
        {
          "regime": "late",
          "epoch": 15,
          "train_loss": 0.003426659619435668,
          "train_accuracy": 1.0,
          "val_loss": 1.9572691441410117,
          "val_accuracy": 0.5833333333333334
        }
      ]
    },
    {
      "regime": "all",
      "trainable_parameters": 11179590,
      "best": {
        "regime": "all",
        "epoch": 13,
        "train_loss": 0.00017177685852705812,
        "train_accuracy": 1.0,
        "val_loss": 0.34318829514086246,
        "val_accuracy": 0.9444444444444444
      },
      "history": [
        {
          "regime": "all",
          "epoch": 1,
          "train_loss": 1.8185815215110779,
          "train_accuracy": 0.2777777777777778,
          "val_loss": 1.3909783628251817,
          "val_accuracy": 0.4166666666666667
        },
        {
          "regime": "all",
          "epoch": 2,
          "train_loss": 0.9497283697128296,
          "train_accuracy": 0.6666666666666666,
          "val_loss": 0.7950038115183512,
          "val_accuracy": 0.75
        },
        {
          "regime": "all",
          "epoch": 3,
          "train_loss": 0.3207702860236168,
          "train_accuracy": 0.9444444444444444,
          "val_loss": 0.566481265756819,
          "val_accuracy": 0.8055555555555556
        },
        {
          "regime": "all",
          "epoch": 4,
          "train_loss": 0.07298391157140334,
          "train_accuracy": 1.0,
          "val_loss": 0.35704346828990513,
          "val_accuracy": 0.8611111111111112
        },
        {
          "regime": "all",
          "epoch": 5,
          "train_loss": 0.02721114596351981,
          "train_accuracy": 1.0,
          "val_loss": 0.3063076308204068,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 6,
          "train_loss": 0.0048724950368826585,
          "train_accuracy": 1.0,
          "val_loss": 0.4545410654197137,
          "val_accuracy": 0.8611111111111112
        },
        {
          "regime": "all",
          "epoch": 7,
          "train_loss": 0.0020719389140140265,
          "train_accuracy": 1.0,
          "val_loss": 0.3854920383956697,
          "val_accuracy": 0.8611111111111112
        },
        {
          "regime": "all",
          "epoch": 8,
          "train_loss": 0.0006057575325636814,
          "train_accuracy": 1.0,
          "val_loss": 0.3490754852278365,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 9,
          "train_loss": 0.0003990160912508145,
          "train_accuracy": 1.0,
          "val_loss": 0.34452139772474766,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 10,
          "train_loss": 0.0003456800671604772,
          "train_accuracy": 1.0,
          "val_loss": 0.3439759831461642,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 11,
          "train_loss": 0.0002850109946545369,
          "train_accuracy": 1.0,
          "val_loss": 0.34235679503116345,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 12,
          "train_loss": 0.000204525165221033,
          "train_accuracy": 1.0,
          "val_loss": 0.34182619531121516,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 13,
          "train_loss": 0.00017177685852705812,
          "train_accuracy": 1.0,
          "val_loss": 0.34318829514086246,
          "val_accuracy": 0.9444444444444444
        },
        {
          "regime": "all",
          "epoch": 14,
          "train_loss": 0.0001473251152977658,
          "train_accuracy": 1.0,
          "val_loss": 0.34607492718431687,
          "val_accuracy": 0.9166666666666666
        },
        {
          "regime": "all",
          "epoch": 15,
          "train_loss": 0.00013161913011572324,
          "train_accuracy": 1.0,
          "val_loss": 0.34792653160790604,
          "val_accuracy": 0.8888888888888888
        }
      ]
    }
  ],
  "selected_test": {
    "loss": 0.16842302017741734,
    "accuracy": 0.9166666666666666,
    "correct": 33,
    "count": 36,
    "confusion_matrix_rows_true_columns_predicted": [
      [
        5,
        1,
        0,
        0,
        0,
        0
      ],
      [
        0,
        6,
        0,
        0,
        0,
        0
      ],
      [
        0,
        0,
        6,
        0,
        0,
        0
      ],
      [
        0,
        0,
        0,
        5,
        1,
        0
      ],
      [
        0,
        0,
        0,
        0,
        6,
        0
      ],
      [
        0,
        0,
        0,
        0,
        1,
        5
      ]
    ]
  },
  "sealed_test_example_manifest_indices": [
    108,
    114,
    120,
    126,
    132,
    138
  ],
  "pretrained_trainval_features": {
    "rows": 108,
    "columns": 512,
    "dtype": "float32",
    "contains_test": false
  },
  "activation_figure": {
    "source_image": "Abyssinian_1.jpg",
    "source_image_sha256": "2533197401eebe9410ea4d063f86c43fbd2666f3e8165a38aca155c0d09c21be",
    "semantics": "computed pretrained activations; not saliency",
    "stages": {
      "layer1": {
        "shape_chw": [
          64,
          56,
          56
        ],
        "selected_channels_by_mean_absolute_activation": [
          50,
          8,
          46,
          63,
          11,
          47,
          39,
          61
        ],
        "display_percentiles_per_channel": [
          1,
          99
        ]
      },
      "layer2": {
        "shape_chw": [
          128,
          28,
          28
        ],
        "selected_channels_by_mean_absolute_activation": [
          41,
          74,
          40,
          36,
          58,
          66,
          106,
          84
        ],
        "display_percentiles_per_channel": [
          1,
          99
        ]
      },
      "layer4": {
        "shape_chw": [
          512,
          7,
          7
        ],
        "selected_channels_by_mean_absolute_activation": [
          167,
          244,
          461,
          10,
          180,
          468,
          46,
          266
        ],
        "display_percentiles_per_channel": [
          1,
          99
        ]
      }
    }
  },
  "artifacts": {
    "selection-manifest.csv": {
      "sha256": "5b5479a311cc2189522c478b4b2dbe014619f31831760ccadf419ddc41619f1e",
      "bytes": 29943
    },
    "training-history.csv": {
      "sha256": "3d767deff74d4bfef694d6343480327d06169fade078aee443ce8f2bf2c6a4be",
      "bytes": 3502
    },
    "sealed-test-predictions.csv": {
      "sha256": "27a40727324f9a9c015983575421c3d785af19456440602a61705fe8231eeb03",
      "bytes": 7271
    },
    "pretrained-trainval-features.npz": {
      "sha256": "62d1c1b07a5f4e22cbbb7dda5f64518c1483fe9b706b4eef2b2c3ceea0b0e45b",
      "bytes": 201717
    },
    "preprocessing-contract.png": {
      "sha256": "af1f941d8312db4ad9b7da828b2617e2d0f683b5dd238349f5772b8e173d8ad4",
      "bytes": 287973
    },
    "resnet18-activations.png": {
      "sha256": "fbd2a011d236274034a759e51e15b8f3441e2ed296fcc56444cf4617bae02aa2",
      "bytes": 225636
    },
    "sealed-test-examples.jpg": {
      "sha256": "d9599e255949f08281a97a5722c65c9c11c23fd85b840a5ef7e1e5ec2eadaa02",
      "bytes": 174704
    },
    "transfer-curves.svg": {
      "sha256": "8e0003610bd67b1057717147f3402ad337ca02ac369228b4b7882722359c831d",
      "bytes": 5332
    }
  },
  "build": {
    "command_template": "python build_transfer_evidence.py --images-dir <official-images> --annotations-dir <official-annotations> --weights <resnet18-f37072fd.pth>",
    "builder_sha256": "78c9a0bbf8768a07c35a22c481ac2b8816c3e1a681caafbda27e5e5db3ccc0cd",
    "python": "3.11.11",
    "platform": "macOS-15.7.7-arm64-arm-64bit",
    "torch": "2.13.0",
    "torchvision": "0.28.0",
    "numpy": "2.4.6",
    "pillow": "12.3.0",
    "threads": 8,
    "deterministic_algorithms": true,
    "elapsed_seconds": 109.09981408296153,
    "numeric_reproducibility_note": "protocol and selection are fixed; floating-point last bits can differ across PyTorch/platform builds",
    "last_figure_refresh": {
      "test_evaluations": 0,
      "python": "3.11.11",
      "torch": "2.13.0",
      "torchvision": "0.28.0",
      "numpy": "2.4.6",
      "pillow": "12.3.0"
    }
  }
}