{
  "schema_version": 2,
  "method": "mllmclip-real-checkpoint-cka",
  "dataset": {
    "name": "POPE adversarial / COCO val2014",
    "train_examples": 320,
    "test_examples": 160
  },
  "checkpoints": {
    "teacher": {
      "model_id": "HuggingFaceTB/SmolVLM2-256M-Video-Instruct",
      "revision": "067788b187b95ebe7b2e040b3e4299e342e5b8fd"
    },
    "student": {
      "model_id": "openai/clip-vit-base-patch32",
      "revision": "3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268"
    }
  },
  "metrics": {
    "seed_results": [
      {
        "seed": 42,
        "baseline": {
          "linear_cka": 0.3728191,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.26125
        },
        "method": {
          "linear_cka": 0.3728191,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.26125
        },
        "model_selection": {
          "label_classes": 1,
          "label_diagnostic_valid": false,
          "test_used_for_selection": false
        }
      },
      {
        "seed": 43,
        "baseline": {
          "linear_cka": 0.3359644,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.26375
        },
        "method": {
          "linear_cka": 0.3310811,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.26
        },
        "model_selection": {
          "label_classes": 1,
          "label_diagnostic_valid": false,
          "test_used_for_selection": false
        }
      },
      {
        "seed": 44,
        "baseline": {
          "linear_cka": 0.3393617,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.2675
        },
        "method": {
          "linear_cka": 0.3358021,
          "knn_accuracy": null,
          "neighbor_overlap_at_5": 0.2575
        },
        "model_selection": {
          "label_classes": 1,
          "label_diagnostic_valid": false,
          "test_used_for_selection": false
        }
      }
    ],
    "aggregate": {
      "baseline_linear_cka": {
        "mean": 0.3493817,
        "std": 0.0203683,
        "ci95_low": 0.3263328,
        "ci95_high": 0.3724307
      },
      "method_linear_cka": {
        "mean": 0.3465675,
        "std": 0.0228568,
        "ci95_low": 0.3207025,
        "ci95_high": 0.3724324
      },
      "baseline_neighbor_overlap_at_5": {
        "mean": 0.2641667,
        "std": 0.0031458,
        "ci95_low": 0.2606069,
        "ci95_high": 0.2677264
      },
      "method_neighbor_overlap_at_5": {
        "mean": 0.2595833,
        "std": 0.0019094,
        "ci95_low": 0.2574226,
        "ci95_high": 0.261744
      }
    }
  },
  "evaluation_protocol": {
    "seeds": [
      42,
      43,
      44
    ],
    "validation_selects_ridge_only": true,
    "heldout_test_used_for_selection": false,
    "tier": "l2_public_dataset",
    "formal_comparison": true,
    "claim_policy": "formal multi-seed comparison"
  },
  "scope": "The POPE label-count diagnostic is degenerate on this subset and is deliberately reported as null; CKA and neighbor overlap show no stable gain.",
  "manifest_ref": "reproduction:mllmclip",
  "provenance": {
    "artifact_path": "docs/reproductions/2608.25575-mllmclip/metrics/pope-checkpoint-a100-seeds42-44.json",
    "dataset_fingerprint": "not recorded in historical artifact",
    "historical_migration": "historical-metrics-v2-2026-08-09",
    "original_code_commit": "not recorded"
  }
}