{
  "evaluator_version": "2.1.0",
  "hash_specification": {
    "algorithm": "SHA-256",
    "canonical_json": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
    "portfolio_canonical_json_sha256_input": "parsed complete portfolio",
    "portfolio_file_sha256_input": "exact source-file bytes",
    "results_canonical_json_sha256_input": "complete sample-id-ordered results array"
  },
  "items": [
    {
      "classes": [
        "building"
      ],
      "config": {
        "confidence_threshold": 0.5,
        "fixed_diagnostic_risk_thresholds": [
          0,
          0.5,
          0.7,
          0.9
        ],
        "match_iou": 0.5
      },
      "false_negatives": [],
      "false_positives": [
        {
          "bbox": [
            20,
            20,
            35,
            35
          ],
          "class": "building",
          "confidence": 0.92,
          "id": "p-hard-fp"
        }
      ],
      "filter": {
        "applied": true,
        "field": "confidence",
        "operator": ">=",
        "threshold": 0.5
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "27595a6d16d96d05e0c8ee326dd03cbf060d68b86b7834bce600ce7d640c3a97",
        "predictions_post_filter_canonical_json_sha256": "66e3b9311b85e2be0d968fbd59359a2d2147eec52034e7cda8bc84cb3e4139ff",
        "predictions_pre_filter_canonical_json_sha256": "66e3b9311b85e2be0d968fbd59359a2d2147eec52034e7cda8bc84cb3e4139ff",
        "references_canonical_json_sha256": "4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:background-test-hard-negative:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:background-test-hard-negative:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "27595a6d16d96d05e0c8ee326dd03cbf060d68b86b7834bce600ce7d640c3a97",
      "matches": [],
      "metadata": {
        "context": "coastal_hard_negative",
        "date": "2025-07-01",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "Oostende",
        "object_size": "empty",
        "occlusion": "none",
        "ood": false,
        "region": "flanders",
        "resolution_m": 0.25,
        "season": "summer",
        "sensor": "synthetic-rgb",
        "source": "fixture-orthophoto",
        "tile_edge": true,
        "urbanity": "urban",
        "vegetation": "moderate"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        {
          "bbox": [
            20,
            20,
            35,
            35
          ],
          "class": "building",
          "confidence": 0.92,
          "id": "p-hard-fp"
        }
      ],
      "predictions_pre_filter": [
        {
          "bbox": [
            20,
            20,
            35,
            35
          ],
          "class": "building",
          "confidence": 0.92,
          "id": "p-hard-fp"
        }
      ],
      "reference_input_field": "references",
      "references": [],
      "sample_id": "background-test-hard-negative",
      "split": "background-test",
      "task": "object_detection"
    },
    {
      "classes": [
        "building"
      ],
      "config": {
        "confidence_threshold": 0.5,
        "fixed_diagnostic_risk_thresholds": [
          0,
          0.5,
          0.7,
          0.9
        ],
        "match_iou": 0.5
      },
      "false_negatives": [],
      "false_positives": [],
      "filter": {
        "applied": true,
        "field": "confidence",
        "operator": ">=",
        "threshold": 0.5
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "0bfe26519b12db547f156905202badca6a5c5a9759f4d6eeb8c0526ddb846974",
        "predictions_post_filter_canonical_json_sha256": "4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945",
        "predictions_pre_filter_canonical_json_sha256": "4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945",
        "references_canonical_json_sha256": "4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:background-test-pure-empty:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:background-test-pure-empty:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "0bfe26519b12db547f156905202badca6a5c5a9759f4d6eeb8c0526ddb846974",
      "matches": [],
      "metadata": {
        "context": "pure_background",
        "date": "2025-02-01",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "Brussel",
        "object_size": "empty",
        "occlusion": "none",
        "ood": false,
        "region": "brussels",
        "resolution_m": 0.25,
        "season": "winter",
        "sensor": "synthetic-rgb",
        "source": "fixture-orthophoto",
        "tile_edge": false,
        "urbanity": "urban",
        "vegetation": "low"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [],
      "predictions_pre_filter": [],
      "reference_input_field": "references",
      "references": [],
      "sample_id": "background-test-pure-empty",
      "split": "background-test",
      "task": "object_detection"
    },
    {
      "classes": [
        "added",
        "removed"
      ],
      "config": {
        "match_iou": 0.5
      },
      "false_negatives": [
        {
          "bbox": [
            10,
            0,
            15,
            5
          ],
          "class": "removed",
          "id": "r2"
        }
      ],
      "false_positives": [
        {
          "bbox": [
            20,
            0,
            25,
            5
          ],
          "class": "removed",
          "confidence": 0.65,
          "id": "p2"
        }
      ],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "324e224883a99770be116a6a74cae2502f95150b963fd6c92342c3111fe367d5",
        "predictions_post_filter_canonical_json_sha256": "beb52d7c3cec145ad733a8adbe95105a6ff838d8f3cf35e335773b1755dc5c43",
        "predictions_pre_filter_canonical_json_sha256": "beb52d7c3cec145ad733a8adbe95105a6ff838d8f3cf35e335773b1755dc5c43",
        "references_canonical_json_sha256": "0018ab1745225a0dcf6d10ad1434a14bbc108f16d0f281e4d730d613941115d3"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:change-test-wallonia-industrial:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:change-test-wallonia-industrial:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "324e224883a99770be116a6a74cae2502f95150b963fd6c92342c3111fe367d5",
      "matches": [
        {
          "class": "added",
          "confidence": 0.85,
          "event": "added",
          "overlap": 1.0,
          "prediction_id": "p1",
          "reference_id": "r1"
        }
      ],
      "metadata": {
        "context": "industrial_change",
        "date": "2024-05-01/2025-05-01",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "Liège",
        "object_size": "mixed",
        "occlusion": "none",
        "ood": false,
        "region": "wallonia",
        "resolution_m": 0.5,
        "season": "multi-date",
        "sensor": "vector",
        "source": "fixture-picc-temporal",
        "tile_edge": false,
        "urbanity": "industrial",
        "vegetation": "low"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        {
          "bbox": [
            0,
            0,
            5,
            5
          ],
          "class": "added",
          "confidence": 0.85,
          "id": "p1"
        },
        {
          "bbox": [
            20,
            0,
            25,
            5
          ],
          "class": "removed",
          "confidence": 0.65,
          "id": "p2"
        }
      ],
      "predictions_pre_filter": [
        {
          "bbox": [
            0,
            0,
            5,
            5
          ],
          "class": "added",
          "confidence": 0.85,
          "id": "p1"
        },
        {
          "bbox": [
            20,
            0,
            25,
            5
          ],
          "class": "removed",
          "confidence": 0.65,
          "id": "p2"
        }
      ],
      "reference_input_field": "references",
      "references": [
        {
          "bbox": [
            0,
            0,
            5,
            5
          ],
          "class": "added",
          "id": "r1"
        },
        {
          "bbox": [
            10,
            0,
            15,
            5
          ],
          "class": "removed",
          "id": "r2"
        }
      ],
      "sample_id": "change-test-wallonia-industrial",
      "split": "test",
      "task": "change_detection"
    },
    {
      "classes": [
        "building"
      ],
      "config": {
        "confidence_threshold": 0.5,
        "fixed_diagnostic_risk_thresholds": [
          0,
          0.5,
          0.7,
          0.9
        ],
        "match_iou": 0.5
      },
      "false_negatives": [],
      "false_positives": [
        {
          "bbox": [
            85,
            85,
            95,
            95
          ],
          "class": "building",
          "confidence": 0.6,
          "id": "p3"
        }
      ],
      "filter": {
        "applied": true,
        "field": "confidence",
        "operator": ">=",
        "threshold": 0.5
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "9bcda4e622a057261be4bbb977064f6e4fd1ae9fa19289124e3ba4de5f829722",
        "predictions_post_filter_canonical_json_sha256": "7c0f9832b4468f46d783b403494378e40a8f21dff3fb5d58545fa3ec06a31d0b",
        "predictions_pre_filter_canonical_json_sha256": "8bee0a2cbd449de9a008b2ed9c9f0b08ca1f7b86b2ed8ac2171de7092089e00c",
        "references_canonical_json_sha256": "238b628e55219d3fbd6e7b38d9ace75f535947e49e83191b8d46233cfcee45ff"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:det-test-flanders-urban:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:det-test-flanders-urban:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "9bcda4e622a057261be4bbb977064f6e4fd1ae9fa19289124e3ba4de5f829722",
      "matches": [
        {
          "class": "building",
          "confidence": 0.9,
          "overlap": 1.0,
          "prediction_id": "p1",
          "reference_id": "r1"
        },
        {
          "class": "building",
          "confidence": 0.7,
          "overlap": 0.8711111111111111,
          "prediction_id": "p2",
          "reference_id": "r2"
        }
      ],
      "metadata": {
        "context": "dense_urban",
        "date": "2025-06-15",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "Antwerpen",
        "object_size": "mixed",
        "occlusion": "partial",
        "ood": false,
        "region": "flanders",
        "resolution_m": 0.25,
        "season": "summer",
        "sensor": "synthetic-rgb",
        "source": "fixture-orthophoto",
        "tile_edge": false,
        "urbanity": "urban",
        "vegetation": "moderate"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        {
          "bbox": [
            10,
            10,
            30,
            30
          ],
          "class": "building",
          "confidence": 0.9,
          "id": "p1"
        },
        {
          "bbox": [
            52,
            52,
            80,
            80
          ],
          "class": "building",
          "confidence": 0.7,
          "id": "p2"
        },
        {
          "bbox": [
            85,
            85,
            95,
            95
          ],
          "class": "building",
          "confidence": 0.6,
          "id": "p3"
        }
      ],
      "predictions_pre_filter": [
        {
          "bbox": [
            10,
            10,
            30,
            30
          ],
          "class": "building",
          "confidence": 0.9,
          "id": "p1"
        },
        {
          "bbox": [
            52,
            52,
            80,
            80
          ],
          "class": "building",
          "confidence": 0.7,
          "id": "p2"
        },
        {
          "bbox": [
            85,
            85,
            95,
            95
          ],
          "class": "building",
          "confidence": 0.6,
          "id": "p3"
        },
        {
          "bbox": [
            2,
            80,
            7,
            86
          ],
          "class": "building",
          "confidence": 0.2,
          "id": "p4-filtered"
        }
      ],
      "reference_input_field": "references",
      "references": [
        {
          "bbox": [
            10,
            10,
            30,
            30
          ],
          "class": "building",
          "id": "r1"
        },
        {
          "bbox": [
            50,
            50,
            80,
            80
          ],
          "class": "building",
          "id": "r2"
        }
      ],
      "sample_id": "det-test-flanders-urban",
      "split": "test",
      "task": "object_detection"
    },
    {
      "classes": [
        0,
        1,
        2
      ],
      "config": {
        "class_order": [
          0,
          1,
          2
        ],
        "masked_pixels_excluded": true
      },
      "false_negatives": [],
      "false_positives": [
        {
          "column": 3,
          "pixel_index": 3,
          "prediction": 2,
          "reason": "class_mismatch",
          "reference": 1,
          "row": 0
        },
        {
          "column": 1,
          "pixel_index": 9,
          "prediction": 1,
          "reason": "class_mismatch",
          "reference": 2,
          "row": 2
        }
      ],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "47640ac47be2509359172b307c164c2888ce4984b82eebf4edd450ec266beae9",
        "predictions_post_filter_canonical_json_sha256": "9dead8589ca1c9f284a7df378a2fe3b817aa78c6d4595c0333cfc08d95f91297",
        "predictions_pre_filter_canonical_json_sha256": "9dead8589ca1c9f284a7df378a2fe3b817aa78c6d4595c0333cfc08d95f91297",
        "references_canonical_json_sha256": "52797b4d5180a470c1e651ea70af03c940a39ee8de82dcb341d23b12cfb1810e"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:raster-test-brussels-urban:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:raster-test-brussels-urban:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "47640ac47be2509359172b307c164c2888ce4984b82eebf4edd450ec266beae9",
      "matches": [],
      "metadata": {
        "context": "dense_urban_raster",
        "date": "2025-10-05",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "Brussel",
        "object_size": "not_applicable",
        "occlusion": "not_applicable",
        "ood": false,
        "region": "brussels",
        "resolution_m": 1,
        "season": "autumn",
        "sensor": "synthetic-multispectral",
        "source": "fixture-thematic-raster",
        "tile_edge": true,
        "urbanity": "urban",
        "vegetation": "moderate"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        [
          0,
          0,
          1,
          2
        ],
        [
          0,
          1,
          1,
          2
        ],
        [
          2,
          1,
          1,
          0
        ]
      ],
      "predictions_pre_filter": [
        [
          0,
          0,
          1,
          2
        ],
        [
          0,
          1,
          1,
          2
        ],
        [
          2,
          1,
          1,
          0
        ]
      ],
      "reference_input_field": "references",
      "references": [
        [
          0,
          0,
          1,
          1
        ],
        [
          0,
          1,
          1,
          2
        ],
        [
          2,
          2,
          1,
          0
        ]
      ],
      "sample_id": "raster-test-brussels-urban",
      "split": "test",
      "task": "raster_classification"
    },
    {
      "classes": [
        "building"
      ],
      "config": {
        "boundary_tolerance_m": 1,
        "match_iou": 0.5
      },
      "false_negatives": [],
      "false_positives": [],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "6c63b5b8c2943b5f393569f8520f031ebfe988b621ea2e6894c6d446928ca6c4",
        "predictions_post_filter_canonical_json_sha256": "dfed6b0f5ac2b7280c7729414a19951b8d074ce00eb83c7d5eaa49e7bdefe135",
        "predictions_pre_filter_canonical_json_sha256": "dfed6b0f5ac2b7280c7729414a19951b8d074ce00eb83c7d5eaa49e7bdefe135",
        "references_canonical_json_sha256": "c5bd8154f877c9d7c5691e6c63d77657ce83ca0a5b1d8fadadccb6ab0a5ff4a9"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:seg-test-wallonia-rural:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:seg-test-wallonia-rural:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "6c63b5b8c2943b5f393569f8520f031ebfe988b621ea2e6894c6d446928ca6c4",
      "matches": [
        {
          "boundary_f1": 1.0,
          "centroid_distance_m": 0.5,
          "class": "building",
          "confidence": 0.8,
          "dice": 0.8881578947368421,
          "overlap": 0.7988165680473372,
          "prediction_id": "p1",
          "reference_id": "r1",
          "relative_area_error": -0.1
        }
      ],
      "metadata": {
        "context": "rural_buildings",
        "date": "2025-04-20",
        "difficulty": "normal",
        "label_review_state": "fixture_reviewed",
        "municipality": "Namur",
        "object_size": "large",
        "occlusion": "none",
        "ood": false,
        "region": "wallonia",
        "resolution_m": 0.5,
        "season": "spring",
        "sensor": "synthetic-rgb",
        "source": "fixture-picc",
        "tile_edge": false,
        "urbanity": "rural",
        "vegetation": "low"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        {
          "class": "building",
          "confidence": 0.8,
          "id": "p1",
          "polygon": [
            [
              0.5,
              0.5
            ],
            [
              9.5,
              0.5
            ],
            [
              9.5,
              8.5
            ],
            [
              0.5,
              8.5
            ],
            [
              0.5,
              0.5
            ]
          ]
        }
      ],
      "predictions_pre_filter": [
        {
          "class": "building",
          "confidence": 0.8,
          "id": "p1",
          "polygon": [
            [
              0.5,
              0.5
            ],
            [
              9.5,
              0.5
            ],
            [
              9.5,
              8.5
            ],
            [
              0.5,
              8.5
            ],
            [
              0.5,
              0.5
            ]
          ]
        }
      ],
      "reference_input_field": "references",
      "references": [
        {
          "class": "building",
          "id": "r1",
          "polygon": [
            [
              0,
              0
            ],
            [
              10,
              0
            ],
            [
              10,
              8
            ],
            [
              0,
              8
            ],
            [
              0,
              0
            ]
          ]
        }
      ],
      "sample_id": "seg-test-wallonia-rural",
      "split": "test",
      "task": "footprint_segmentation"
    },
    {
      "classes": null,
      "config": {
        "missing_prediction_policy": "exclude_and_reduce_coverage"
      },
      "false_negatives": [
        {
          "missing_index": 3
        }
      ],
      "false_positives": [],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "d9e2637e15491546f8eaad2dc140fe2af78f213dd9899bc0f683234880a5e0c8",
        "predictions_post_filter_canonical_json_sha256": "74d1fc53e85cc8ba6ca6cd8fb81a998ada3e3dd0e7d8ef5967f9ec8cad722768",
        "predictions_pre_filter_canonical_json_sha256": "74d1fc53e85cc8ba6ca6cd8fb81a998ada3e3dd0e7d8ef5967f9ec8cad722768",
        "references_canonical_json_sha256": "592601848bf07aeef289349574a01b11c006e801f47a29ddd584113b146f7390"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:terrain-test-flanders-rural:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:terrain-test-flanders-rural:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "d9e2637e15491546f8eaad2dc140fe2af78f213dd9899bc0f683234880a5e0c8",
      "matches": [],
      "metadata": {
        "context": "terrain_profile",
        "date": "2023-01-01",
        "difficulty": "normal",
        "label_review_state": "fixture_reviewed",
        "municipality": "Hasselt",
        "object_size": "not_applicable",
        "occlusion": "not_applicable",
        "ood": false,
        "region": "flanders",
        "resolution_m": 5,
        "season": "not_applicable",
        "sensor": "elevation-raster",
        "source": "fixture-dhmv",
        "tile_edge": false,
        "urbanity": "rural",
        "vegetation": "moderate"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        12.2,
        12.8,
        14.4,
        null,
        15.7
      ],
      "predictions_pre_filter": [
        12.2,
        12.8,
        14.4,
        null,
        15.7
      ],
      "reference_input_field": "references",
      "references": [
        12,
        13,
        14,
        15,
        16
      ],
      "sample_id": "terrain-test-flanders-rural",
      "split": "test",
      "task": "terrain_interpretation"
    },
    {
      "classes": null,
      "config": {
        "critical_severities": [
          "blocker",
          "critical"
        ]
      },
      "false_negatives": [
        {
          "anomaly": "D-MISSING-PROVENANCE",
          "severity": "critical"
        }
      ],
      "false_positives": [
        {
          "anomaly": "D-RESOLUTION",
          "severity": "major"
        }
      ],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "cd191708e28435027749f573412592dd7ce2c4cd1e08656f6306382691a9d2fb",
        "predictions_post_filter_canonical_json_sha256": "a9f7113f70c1b41d8927ba6018a8e22fbeb15499f0833ad16dc9be3b1d567371",
        "predictions_pre_filter_canonical_json_sha256": "a9f7113f70c1b41d8927ba6018a8e22fbeb15499f0833ad16dc9be3b1d567371",
        "references_canonical_json_sha256": "1d19e53c82b4815685a23921146ce35c0105f403774b712c1c41ee3ca4aa0672"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:validation-test-national:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:validation-test-national:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "cd191708e28435027749f573412592dd7ce2c4cd1e08656f6306382691a9d2fb",
      "matches": [
        {
          "anomaly": "D-CRS-WRONG",
          "expected_severity": "blocker",
          "observed_severity": "blocker"
        },
        {
          "anomaly": "D-INVALID-GEOMETRY",
          "expected_severity": "critical",
          "observed_severity": "critical"
        }
      ],
      "metadata": {
        "context": "data_contract",
        "date": "2025-01-01",
        "difficulty": "hard",
        "label_review_state": "fixture_reviewed",
        "municipality": "not_applicable",
        "object_size": "not_applicable",
        "occlusion": "not_applicable",
        "ood": false,
        "region": "national",
        "resolution_m": null,
        "season": "not_applicable",
        "sensor": "mixed",
        "source": "fixture-contracts",
        "tile_edge": false,
        "urbanity": "mixed",
        "vegetation": "mixed"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "observed_anomalies",
      "predictions_post_filter": [
        {
          "code": "D-CRS-WRONG",
          "severity": "blocker"
        },
        {
          "code": "D-INVALID-GEOMETRY",
          "severity": "critical"
        },
        {
          "code": "D-RESOLUTION",
          "severity": "major"
        }
      ],
      "predictions_pre_filter": [
        {
          "code": "D-CRS-WRONG",
          "severity": "blocker"
        },
        {
          "code": "D-INVALID-GEOMETRY",
          "severity": "critical"
        },
        {
          "code": "D-RESOLUTION",
          "severity": "major"
        }
      ],
      "reference_input_field": "expected_anomalies",
      "references": [
        {
          "code": "D-CRS-WRONG",
          "severity": "blocker"
        },
        {
          "code": "D-INVALID-GEOMETRY",
          "severity": "critical"
        },
        {
          "code": "D-MISSING-PROVENANCE",
          "severity": "critical"
        }
      ],
      "sample_id": "validation-test-national",
      "split": "test",
      "task": "geospatial_data_validation"
    },
    {
      "classes": [
        "building"
      ],
      "config": {
        "match_iou": 0.5
      },
      "false_negatives": [
        {
          "class": "building",
          "id": "r2",
          "polygon": [
            [
              10,
              0
            ],
            [
              14,
              0
            ],
            [
              14,
              4
            ],
            [
              10,
              4
            ],
            [
              10,
              0
            ]
          ]
        }
      ],
      "false_positives": [
        {
          "class": "building",
          "id": "p2",
          "polygon": [
            [
              20,
              0
            ],
            [
              24,
              0
            ],
            [
              24,
              4
            ],
            [
              20,
              4
            ],
            [
              20,
              0
            ]
          ]
        }
      ],
      "filter": {
        "applied": false
      },
      "hashes": {
        "canonicalization": "UTF-8 JSON produced with sort_keys=true, separators=(',', ':'), ensure_ascii=false and allow_nan=false",
        "case_input_canonical_json_sha256": "e3b554b9b91a89214b9ed7703f4fb3463607024f35a33b934c94ce8c783d405a",
        "predictions_post_filter_canonical_json_sha256": "b2a08e0928cd1d93e4548f95a1a40bdf404891bdaa9b876cce0f684edcce3fed",
        "predictions_pre_filter_canonical_json_sha256": "b2a08e0928cd1d93e4548f95a1a40bdf404891bdaa9b876cce0f684edcce3fed",
        "references_canonical_json_sha256": "06b0c3de29764b1ccc34e1094e7d5aaadace27fa8f4dc4d6022f01db441ad663"
      },
      "input_lineage": {
        "prediction": {
          "derivation": "hand_authored_fixed_output",
          "source_id": "synthetic:vector-test-flanders-suburban:prediction",
          "source_version": "1"
        },
        "reference": {
          "derivation": "hand_authored_contract_fixture",
          "source_id": "synthetic:vector-test-flanders-suburban:reference",
          "source_version": "1"
        }
      },
      "input_sha256": "e3b554b9b91a89214b9ed7703f4fb3463607024f35a33b934c94ce8c783d405a",
      "matches": [
        {
          "class": "building",
          "confidence": null,
          "overlap": 1.0,
          "prediction_id": "p1",
          "reference_id": "r1"
        }
      ],
      "metadata": {
        "context": "suburban_buildings",
        "date": "2025-01-15",
        "difficulty": "normal",
        "label_review_state": "fixture_reviewed",
        "municipality": "Mol",
        "object_size": "mixed",
        "occlusion": "not_applicable",
        "ood": false,
        "region": "flanders",
        "resolution_m": 0.1,
        "season": "not_applicable",
        "sensor": "vector",
        "source": "fixture-grb",
        "tile_edge": false,
        "urbanity": "suburban",
        "vegetation": "low"
      },
      "portfolio_lineage": {
        "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.",
        "declared": {
          "origin": "repository_fixture",
          "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
          "version": "1"
        },
        "portfolio_canonical_json_sha256": "1ac1bd77db125db48f9168406d51bfb723522f54a6c81dbe868d600c72f27b17",
        "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
        "portfolio_id": "geointel-p4-reference-harness-v2",
        "portfolio_kind": "synthetic_contract",
        "portfolio_schema_version": 2,
        "source_path": "fixtures/accuracy/p4/protected-baseline-cases.json",
        "split_roles": [
          "test",
          "background-test"
        ]
      },
      "prediction_input_field": "predictions",
      "predictions_post_filter": [
        {
          "class": "building",
          "id": "p1",
          "polygon": [
            [
              0,
              0
            ],
            [
              4,
              0
            ],
            [
              4,
              4
            ],
            [
              0,
              4
            ],
            [
              0,
              0
            ]
          ]
        },
        {
          "class": "building",
          "id": "p2",
          "polygon": [
            [
              20,
              0
            ],
            [
              24,
              0
            ],
            [
              24,
              4
            ],
            [
              20,
              4
            ],
            [
              20,
              0
            ]
          ]
        }
      ],
      "predictions_pre_filter": [
        {
          "class": "building",
          "id": "p1",
          "polygon": [
            [
              0,
              0
            ],
            [
              4,
              0
            ],
            [
              4,
              4
            ],
            [
              0,
              4
            ],
            [
              0,
              0
            ]
          ]
        },
        {
          "class": "building",
          "id": "p2",
          "polygon": [
            [
              20,
              0
            ],
            [
              24,
              0
            ],
            [
              24,
              4
            ],
            [
              20,
              4
            ],
            [
              20,
              0
            ]
          ]
        }
      ],
      "reference_input_field": "references",
      "references": [
        {
          "class": "building",
          "id": "r1",
          "polygon": [
            [
              0,
              0
            ],
            [
              4,
              0
            ],
            [
              4,
              4
            ],
            [
              0,
              4
            ],
            [
              0,
              0
            ]
          ]
        },
        {
          "class": "building",
          "id": "r2",
          "polygon": [
            [
              10,
              0
            ],
            [
              14,
              0
            ],
            [
              14,
              4
            ],
            [
              10,
              4
            ],
            [
              10,
              0
            ]
          ]
        }
      ],
      "sample_id": "vector-test-flanders-suburban",
      "split": "test",
      "task": "vector_comparison"
    }
  ],
  "items_canonical_json_sha256": "14c27341951d8d250578013d7ae3c5fe6c585a6ea84752410ab786fc2d5f805a",
  "portfolio_file_sha256": "c49a9c8d74cae2062552b012d9c275f65156c81d75b3e453860f5b859c1c5473",
  "schema_version": 2
}