{ "benchmark_id": "geointel-p4-reference-harness-v2", "claim_boundary": "Synthetic deterministic reference cases validate evaluator behavior, not production model accuracy.", "evaluator_families": [ "change_detection", "footprint_segmentation", "geospatial_data_validation", "object_detection", "raster_classification", "terrain_interpretation", "vector_comparison" ], "evaluator_version": "2.1.0", "gate_states": [ "fail", "not_evaluable", "pass" ], "implemented_capabilities": [ { "capability_id": "model_object_detection", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "model_inference_pipeline", "implementation_paths": [ "backend/app/services/detection_service.py" ], "suitable_metrics": [ "precision", "recall", "F1", "AP50", "AP50-95", "IoU", "ECE", "Brier", "coverage-risk" ], "task": "object_detection" }, { "capability_id": "building_proposal_filtering", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "not_separately_benchmarked", "implementation_kind": "supporting_candidate_classifier", "implementation_paths": [ "scripts/train_building_proposal_classifier.py", "scripts/evaluate_belgium_building_candidate.py" ], "suitable_metrics": [ "candidate precision", "candidate recall", "F1", "calibration" ], "task": "object_detection" }, { "capability_id": "footprint_segmentation", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "model_inference_pipeline", "implementation_paths": [ "backend/app/services/segmentation_service.py", "backend/app/services/segmentation_adapter.py" ], "suitable_metrics": [ "object precision", "object recall", "F1", "IoU", "Dice", "boundary F1", "centroid distance", "area error", "topology" ], "task": "footprint_segmentation" }, { "capability_id": "detection_and_vector_qa", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "deterministic_geospatial_comparison", "implementation_paths": [ "backend/app/services/qa_service.py", "backend/app/services/detection_qa_service.py", "backend/app/services/quality_check_service.py" ], "suitable_metrics": [ "precision", "recall", "F1", "IoU", "topology", "coverage" ], "task": "vector_comparison" }, { "capability_id": "vector_clip_buffer_intersect", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "covered_by_family_not_separately_benchmarked", "implementation_kind": "deterministic_vector_processing", "implementation_paths": [ "backend/app/services/vector_operations_service.py", "backend/app/services/vector_feature_service.py" ], "suitable_metrics": [ "geometry validity", "CRS correctness", "area conservation", "feature counts", "topology" ], "task": "vector_comparison" }, { "capability_id": "temporal_vector_change", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "deterministic_change_detection", "implementation_paths": [ "backend/app/services/change_detection_service.py" ], "suitable_metrics": [ "event precision", "event recall", "event F1", "IoU" ], "task": "change_detection" }, { "capability_id": "thematic_raster_interpretation", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_metric_contract_only_no_generic_learned_classifier_claim", "implementation_kind": "deterministic_source_interpretation", "implementation_paths": [ "backend/app/services/thematic_raster_analysis_service.py" ], "suitable_metrics": [ "pixel accuracy", "per-class F1", "per-class IoU", "mean IoU" ], "task": "raster_classification" }, { "capability_id": "raster_clip_reproject_indices", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "covered_by_family_not_separately_benchmarked", "implementation_kind": "deterministic_raster_processing", "implementation_paths": [ "backend/app/services/raster_service.py", "backend/app/services/raster_operations_service.py" ], "suitable_metrics": [ "CRS/transform preservation", "pixel alignment", "nodata", "numeric tolerance" ], "task": "raster_classification" }, { "capability_id": "raster_partition_mosaic", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "not_separately_benchmarked", "implementation_kind": "deterministic_raster_partitioning", "implementation_paths": [ "backend/app/services/raster_partition_analysis_service.py" ], "suitable_metrics": [ "seam equality", "coverage completeness", "resolution consistency" ], "task": "raster_classification" }, { "capability_id": "terrain_height_interpretation", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "deterministic_continuous_raster_analysis", "implementation_paths": [ "backend/app/services/terrain_analysis_service.py", "backend/app/services/spw_terrain_service.py" ], "suitable_metrics": [ "MAE", "RMSE", "bias", "coverage", "unit integrity" ], "task": "terrain_interpretation" }, { "capability_id": "flood_hazard_interpretation", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "covered_by_family_not_separately_benchmarked", "implementation_kind": "deterministic_scenario_raster_analysis", "implementation_paths": [ "backend/app/services/flood_hazard_analysis_service.py" ], "suitable_metrics": [ "depth MAE/RMSE", "hazard-class IoU", "coverage", "scenario identity" ], "task": "terrain_interpretation" }, { "capability_id": "bathymetry_interpretation", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "covered_by_family_not_separately_benchmarked", "implementation_kind": "deterministic_vertical_reference_analysis", "implementation_paths": [ "backend/app/services/bathymetry_raster_analysis_service.py", "backend/app/services/mdk_bathymetry_probe_service.py" ], "suitable_metrics": [ "MAE", "RMSE", "bias", "coverage", "vertical-datum integrity" ], "task": "terrain_interpretation" }, { "capability_id": "data_contract_validation_and_scan", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "synthetic_contract_case_only", "implementation_kind": "deterministic_validation", "implementation_paths": [ "backend/app/services/data_contract_validation.py", "scripts/run_accuracy_phase3_full_data_scan.py" ], "suitable_metrics": [ "anomaly precision", "anomaly recall", "anomaly F1", "critical misses" ], "task": "geospatial_data_validation" }, { "capability_id": "aoi_partition_orchestration", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "not_separately_benchmarked", "implementation_kind": "deterministic_orchestration", "implementation_paths": [ "backend/app/services/aoi_operation_service.py", "backend/app/services/aoi_operation_executor.py", "backend/app/services/aoi_operation_worker.py" ], "suitable_metrics": [ "partition completeness", "overlap/gap", "idempotency", "resume correctness" ], "task": "geospatial_data_validation" }, { "capability_id": "geo_assistant_orchestration", "claim_boundary": "A family mapping is not evidence that this capability has a separate representative product benchmark.", "evaluation_status": "no_independent_accuracy_score_underlying_tool_results_are_authoritative", "implementation_kind": "tool_orchestration_interface", "implementation_paths": [ "backend/app/services/geo_assistant_service.py" ], "suitable_metrics": [ "tool-selection correctness", "grounding", "unsupported-claim rate" ], "task": "geospatial_data_validation" } ], "metric_contract": { "path": "docs/accuracy-program/05-metric-framework.md", "sha256": "18abfc21964b7d98848d8ae513afce242184980c9517c9184d40efd9165ea00d", "size_bytes": 28053 }, "protected_policy": { "diagnostic_curves_select_operating_point": false, "operating_point_selection_allowed": false, "test_feedback_allowed": false, "threshold_selection_source": "pre_registered_configuration_only" }, "schema_version": 2, "split_generator_version": "1.3.0", "undefined_value_policy": "Undefined denominators are null with numerator, denominator and support; they are never coerced to a perfect score.", "workflow_version": "2.0.1" }