{
  "schema_version": 1,
  "adapter_key": "oob-regularization",
  "validated_at": "2026-10-16",
  "accelerator": {
    "vendor": "NVIDIA",
    "model": "NVIDIA A100-SXM4-80GB"
  },
  "command": [
    "python",
    "scripts/run_oob_regularization.py",
    "--dataset-dir",
    "public/cifar10",
    "--device",
    "cuda",
    "--seed",
    "42",
    "--output",
    "runs/oob-42.json"
  ],
  "seeds": [
    42,
    43,
    44
  ],
  "dataset": {
    "id": "CIFAR-10",
    "revision": "sha256:54636561a3ce25bd3e19253c6b0d8538147b0ae398331ac4a2d86c6d987368cd",
    "source_url": "https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz",
    "source_md5": "c99cafc152244af753f735de768cd75f",
    "split_revision": "6ba2882738809786a921455ef0f835759c1cd01d53976deb458fa9d694b1b23e",
    "training_examples": 512,
    "validation_examples": 128,
    "official_test_read": false
  },
  "command_path_policy": "public input paths relativized; exact content revision retained",
  "checkpoint": {
    "public_id": "none-independent-seeded-initialization",
    "revision": "63ca47dad9a19a302b31b1699267dd5c53692b9d",
    "pretrained": false,
    "initialization_seeds": [
      42,
      43,
      44
    ]
  },
  "result": "passed",
  "metrics": {
    "sigreg_in_batch_validation_knn5_accuracy": {
      "mean": 0.20572916666666666,
      "std": 0.06825699178167269,
      "n": 3
    },
    "sigreg_oob_validation_knn5_accuracy": {
      "mean": 0.2265625,
      "std": 0.020669932117692115,
      "n": 3
    },
    "simclr_validation_knn5_accuracy": {
      "mean": 0.21354166666666666,
      "std": 0.009021097956087902,
      "n": 3
    },
    "contrastive_oob_validation_knn5_accuracy": {
      "mean": 0.234375,
      "std": 0.013531646934131853,
      "n": 3
    },
    "sigreg_oob_minus_in_batch": {
      "mean": 0.020833333333333332,
      "std": 0.05083130545818525,
      "n": 3
    },
    "contrastive_oob_minus_simclr": {
      "mean": 0.020833333333333332,
      "std": 0.022552744890219755,
      "n": 3
    }
  },
  "execution_checks": {
    "four_actual_training_arms": true,
    "matched_initialization_and_augmentation_streams": true,
    "cpu_detached_sample_anchor_cache": true,
    "four_refreshes_each_oob_arm": true,
    "validation_not_used_for_selection": true,
    "official_test_read": false,
    "unified_cli_cuda_same_seed_agreement": true
  },
  "diagnostic_only": true,
  "promotion_eligible": false,
  "executed_source_sha256": {
    "src/auto_research/reproductions/oob_regularization/algorithm.py": "283c6b540b04b4bee6ebe7e879a25dc4804973f42c0ee66c9b652aea62f61ad9",
    "src/auto_research/reproductions/oob_regularization/experiment.py": "7a03e24e8028025fd5623aeff39bafa50dcf4f2ad72ffe0b66dd9c414f966ae2",
    "scripts/run_oob_regularization.py": "ac5678c554575608f5dc58d47efa498383c2408008c5c03c5f8dd9eb562c3a0a",
    "src/auto_research/runtime.py": "0a924b7c06e0129f74e8b94ab98b56b05bf9d2e66d8c7d745fc24a66150670e3",
    "src/auto_research/cli.py": "54bbbf251b797580d76022f6baa276415e9f50da6244d37ba157c751cbaa4aaa",
    "src/auto_research/reproductions/reporting.py": "ccf2f6de0d3c2cb22d5c21151bc9feecd3769ff5a8f83fbd3aafec395233adb7"
  },
  "prior_failed_attempts": [
    {
      "commit": "d63fa6f67bb1d0e052cf0eabf8254e383697e6b7",
      "error_type": "TypeError",
      "reason": "unified CLI adapter used unsupported device_for keyword arguments",
      "resolution": "shared runtime API and explicit CUDA index context; regression tests and real unified CLI rerun"
    }
  ],
  "provenance": {
    "commit": "63ca47dad9a19a302b31b1699267dd5c53692b9d",
    "artifact_path": "docs/gpu-validations/oob-regularization-a100-20261016.json",
    "metric_artifact_path": "docs/reproductions/2610.20158-oob-regularization/metrics/component-seeds42-44.json"
  },
  "full_experiment_status": "not_reproduced_high_cost",
  "additional_validated_commands": [
    [
      "python",
      "-m",
      "auto_research.cli",
      "reproduce",
      "--paper",
      "oob-regularization",
      "--device",
      "cuda",
      "--cpu-threads",
      "1",
      "--seeds",
      "42",
      "--budget",
      "smoke",
      "--dataset-dir",
      "public/cifar10",
      "--output-dir",
      "runs/oob-cli"
    ]
  ]
}
