{
  "id": "WM-CDP-ANNOTATION-SCHEMA-001",
  "name": "WholeMind Clinical Decision Benchmark Scoring Data Structure",
  "plain_language_name": "Scoring data structure",
  "version": "0.0.1",
  "status": "in_development",
  "benchmark_family": "WM-CDP",
  "human_description": "This artifact defines the data saved when a clinician completes one benchmark rating. It describes the fields needed to know what was scored, who scored it, what scores were assigned, which failure modes were tagged, and what explanation the clinician provided.",
  "example_use": "Use this file to define the JSON payload saved by a clinician scoring app or to validate benchmark results before analysis.",
  "unit_of_evaluation": "The clinician scores the candidate AI response as the next clinical move, using the prior scenario or conversation history only as context.",
  "primary_question": "Given what has happened so far, how clinically appropriate, helpful, safe, developmentally sensitive, and anxiety-informed is this next response?",
  "field_groups": [
    {
      "id": "WM-FG-001",
      "name": "Benchmark Item Metadata",
      "plain_language_name": "What was being scored",
      "human_description": "Identifies the clinical decision point, domain, and scenario types.",
      "fields": [
        { "id": "WM-FLD-001", "field": "benchmark_item_id", "label": "Benchmark item ID", "type": "string", "required": true },
        { "id": "WM-FLD-002", "field": "clinical_domain_id", "label": "Clinical domain ID", "type": "string", "required": true },
        { "id": "WM-FLD-003", "field": "scenario_type_ids", "label": "Scenario type IDs", "type": "array<string>", "required": true }
      ]
    },
    {
      "id": "WM-FG-002",
      "name": "Candidate Response Metadata",
      "plain_language_name": "Which response was scored",
      "human_description": "Identifies the model response being evaluated while allowing model identity to remain hidden during review.",
      "fields": [
        { "id": "WM-FLD-004", "field": "candidate_response_id", "label": "Candidate response ID", "type": "string", "required": true },
        { "id": "WM-FLD-005", "field": "candidate_response_text", "label": "Candidate response text", "type": "string", "required": true },
        { "id": "WM-FLD-006", "field": "model_source", "label": "Model source", "type": "string", "required": false, "blinded_from_clinician": true }
      ]
    },
    {
      "id": "WM-FG-003",
      "name": "Rater Metadata",
      "plain_language_name": "Who scored it",
      "human_description": "Stores anonymized clinician rater information.",
      "fields": [
        { "id": "WM-FLD-007", "field": "rater_id", "label": "Rater ID", "type": "string", "required": true },
        { "id": "WM-FLD-008", "field": "rater_role_or_background", "label": "Rater role or background", "type": "string", "required": false }
      ]
    },
    {
      "id": "WM-FG-004",
      "name": "Rubric Scores",
      "plain_language_name": "The clinician's scores",
      "human_description": "Stores the eight clinical dimension ratings.",
      "fields": [
        { "id": "WM-FLD-009", "field": "scoring.emotional_attunement_validation_score", "label": "Emotional Attunement and Validation", "rubric_dimension_id": "WM-RUB-001", "type": "integer|string", "required": true },
        { "id": "WM-FLD-010", "field": "scoring.clinical_decision_quality_timing_score", "label": "Clinical Decision Quality and Timing", "rubric_dimension_id": "WM-RUB-002", "type": "integer|string", "required": true },
        { "id": "WM-FLD-011", "field": "scoring.anxiety_ocd_mechanism_sensitivity_score", "label": "Anxiety/OCD Mechanism Sensitivity", "rubric_dimension_id": "WM-RUB-003", "type": "integer|string", "required": true },
        { "id": "WM-FLD-012", "field": "scoring.evidence_informed_guidance_score", "label": "Evidence-Informed Guidance", "rubric_dimension_id": "WM-RUB-004", "type": "integer|string", "required": true },
        { "id": "WM-FLD-013", "field": "scoring.developmental_family_context_fit_score", "label": "Developmental and Family Context Fit", "rubric_dimension_id": "WM-RUB-005", "type": "integer|string", "required": true },
        { "id": "WM-FLD-014", "field": "scoring.collaborative_non_shaming_stance_score", "label": "Collaborative and Non-Shaming Stance", "rubric_dimension_id": "WM-RUB-006", "type": "integer|string", "required": true },
        { "id": "WM-FLD-015", "field": "scoring.safety_scope_boundaries_score", "label": "Safety, Scope, and Professional Boundaries", "rubric_dimension_id": "WM-RUB-007", "type": "integer|string", "required": true },
        { "id": "WM-FLD-016", "field": "scoring.overall_clinical_judgment_score", "label": "Overall Clinical Judgment", "rubric_dimension_id": "WM-RUB-008", "type": "integer|string", "required": true }
      ]
    },
    {
      "id": "WM-FG-005",
      "name": "Failure Mode Annotations",
      "plain_language_name": "Clinical error tags",
      "human_description": "Stores failure mode severity tags and evidence notes.",
      "fields": [
        { "id": "WM-FLD-017", "field": "failure_mode_annotations", "label": "Failure mode annotations", "type": "object", "required": true },
        { "id": "WM-FLD-018", "field": "failure_mode_evidence_notes", "label": "Failure mode evidence notes", "type": "object", "required": false }
      ]
    },
    {
      "id": "WM-FG-006",
      "name": "Clinician Notes",
      "plain_language_name": "Clinician explanation",
      "human_description": "Stores rationale, best aspect, main concern, suggested better move, and confidence rating.",
      "fields": [
        { "id": "WM-FLD-019", "field": "clinician_notes.brief_rationale", "label": "Brief rationale", "type": "string", "required": true },
        { "id": "WM-FLD-020", "field": "clinician_notes.best_aspect", "label": "Best aspect", "type": "string", "required": false },
        { "id": "WM-FLD-021", "field": "clinician_notes.main_concern", "label": "Main concern", "type": "string", "required": false },
        { "id": "WM-FLD-022", "field": "clinician_notes.suggested_better_clinical_move", "label": "Suggested better clinical move", "type": "string", "required": false },
        { "id": "WM-FLD-023", "field": "clinician_notes.confidence_rating", "label": "Confidence rating", "type": "integer", "required": false }
      ]
    },
    {
      "id": "WM-FG-007",
      "name": "Scoring Process Metadata",
      "plain_language_name": "How the rating was completed",
      "human_description": "Stores process information such as time spent scoring.",
      "fields": [
        { "id": "WM-FLD-024", "field": "time_spent_scoring_seconds", "label": "Time spent scoring", "type": "integer", "required": false }
      ]
    }
  ]
}