data/evaluation/aggregate_l3b.json

{
  "format_version": 1,
  "config_fingerprint": "v06_l3b_retro_archive",
  "evaluated_at": "2026-06-07T17:05:34.785249Z",
  "n_events": 6,
  "n_seeds": 3,
  "seed_list": [
    42,
    1337,
    2026
  ],
  "bootstrap": {
    "n_resamples": 1000,
    "seed": 42,
    "confidence": 0.95
  },
  "metrics": {
    "category_recall": {
      "mean": 0.45630511463844803,
      "ci_low": 0.31261022927689597,
      "ci_high": 0.6199140211640212,
      "n": 6,
      "n_resamples": 1000,
      "confidence": 0.95,
      "per_event_values": {
        "2024-0077-GRC": 0.5,
        "2024-0183-FRA": 0.25,
        "2024-0796-ESP": 0.25925925925925924,
        "2025-0183-ITA": 0.42857142857142855,
        "2025-0848-UKR": 0.5,
        "2025-0852-BGR": 0.8000000000000002
      },
      "per_event_seed_values": {
        "2024-0077-GRC": [
          0.5,
          0.5,
          0.5
        ],
        "2024-0183-FRA": [
          0.25,
          0.25,
          0.25
        ],
        "2024-0796-ESP": [
          0.25925925925925924,
          0.25925925925925924,
          0.25925925925925924
        ],
        "2025-0183-ITA": [
          0.42857142857142855,
          0.42857142857142855,
          0.42857142857142855
        ],
        "2025-0848-UKR": [
          0.5,
          0.5,
          0.5
        ],
        "2025-0852-BGR": [
          0.8,
          0.8,
          0.8
        ]
      }
    },
    "category_severity_match_rate": {
      "mean": 0.6706349206349206,
      "ci_low": 0.5277777777777778,
      "ci_high": 0.8374007936507937,
      "n": 6,
      "n_resamples": 1000,
      "confidence": 0.95,
      "per_event_values": {
        "2024-0077-GRC": 0.5,
        "2024-0183-FRA": 0.5,
        "2024-0796-ESP": 0.8571428571428571,
        "2025-0183-ITA": 1.0,
        "2025-0848-UKR": 0.6666666666666666,
        "2025-0852-BGR": 0.5
      },
      "per_event_seed_values": {
        "2024-0077-GRC": [
          0.5,
          0.5,
          0.5
        ],
        "2024-0183-FRA": [
          0.5,
          0.5,
          0.5
        ],
        "2024-0796-ESP": [
          0.8571428571428571,
          0.8571428571428571,
          0.8571428571428571
        ],
        "2025-0183-ITA": [
          1.0,
          1.0,
          1.0
        ],
        "2025-0848-UKR": [
          0.6666666666666666,
          0.6666666666666666,
          0.6666666666666666
        ],
        "2025-0852-BGR": [
          0.5,
          0.5,
          0.5
        ]
      }
    },
    "cosine_f1": {
      "mean": 0.2654166666666667,
      "ci_low": 0.17522500000000002,
      "ci_high": 0.3878183333333333,
      "n": 6,
      "n_resamples": 1000,
      "confidence": 0.95,
      "per_event_values": {
        "2024-0077-GRC": 0.1818,
        "2024-0183-FRA": 0.2857,
        "2024-0796-ESP": 0.2941,
        "2025-0183-ITA": 0.08700000000000001,
        "2025-0848-UKR": 0.5217,
        "2025-0852-BGR": 0.22220000000000004
      },
      "per_event_seed_values": {
        "2024-0077-GRC": [
          0.1818,
          0.1818,
          0.1818
        ],
        "2024-0183-FRA": [
          0.2857,
          0.2857,
          0.2857
        ],
        "2024-0796-ESP": [
          0.2941,
          0.2941,
          0.2941
        ],
        "2025-0183-ITA": [
          0.087,
          0.087,
          0.087
        ],
        "2025-0848-UKR": [
          0.5217,
          0.5217,
          0.5217
        ],
        "2025-0852-BGR": [
          0.2222,
          0.2222,
          0.2222
        ]
      }
    },
    "category_f1": {
      "mean": 0.3813614158882701,
      "ci_low": 0.31142812856367846,
      "ci_high": 0.4538694704935114,
      "n": 6,
      "n_resamples": 1000,
      "confidence": 0.95,
      "per_event_values": {
        "2024-0077-GRC": 0.3636363636363636,
        "2024-0183-FRA": 0.28571428571428575,
        "2024-0796-ESP": 0.411764705882353,
        "2025-0183-ITA": 0.26086956521739124,
        "2025-0848-UKR": 0.5217391304347826,
        "2025-0852-BGR": 0.4444444444444444
      },
      "per_event_seed_values": {
        "2024-0077-GRC": [
          0.36363636363636365,
          0.36363636363636365,
          0.36363636363636365
        ],
        "2024-0183-FRA": [
          0.28571428571428575,
          0.28571428571428575,
          0.28571428571428575
        ],
        "2024-0796-ESP": [
          0.4117647058823529,
          0.4117647058823529,
          0.4117647058823529
        ],
        "2025-0183-ITA": [
          0.26086956521739124,
          0.26086956521739124,
          0.26086956521739124
        ],
        "2025-0848-UKR": [
          0.5217391304347826,
          0.5217391304347826,
          0.5217391304347826
        ],
        "2025-0852-BGR": [
          0.4444444444444444,
          0.4444444444444444,
          0.4444444444444444
        ]
      }
    }
  },
  "source_archive": "data/evaluation/gold/_v10_l3b_authoritative"
}