med-semantic-grading

name: med-semantic-grading description: 基于研究标准的语义分级技能。Use when applying a second-pass semantic grade to triaged items using configurable research focus term lists from workflow_rules.json, after rule-based grading is complete.

med-semantic-grading

1. Purpose:

• Apply a second-pass semantic grade to items already classified by rule-based grading (see med-daily-triage)
• Uses configurable term lists from workflow_rules.json → research_focus field
• Skill is domain-agnostic: all research-specific terms come from user config

2. Data source:

• workflow_rules.json → triage.research_focus field contains three configurable term groups:
  • group1 (config key: core_exposure_terms) — first term group
  • group2 (config key: core_biology_terms) — second term group
  • group3 (config key: mechanism_terms) — third term group
• All three groups are user-defined; skill does not hardcode any domain-specific vocabulary

3. Implementation entry point:

• workflow_classifier.mjs → deriveSemanticGradeFromStandards()
• Called from run_research_os_pipeline.mjs after rule-based triage

4. Scoring logic:

• For each eligible item (B/C only), count term hits in title + abstract:
  • group1Hits: number of group1 terms found in text
  • group2Hits: number of group2 terms found in text
  • group3Hits: number of group3 terms found in text
• Supplement with weak evidence from semantic_search results:
  • searchEvidence: +2 if a search result title contains both group1 and group2 terms; +1 if only group2
• Upgrade rules (only upgrade, never downgrade for missing signals):
  • strongUpgrade: (group1 >= 1 AND group2 >= 2) OR (group2 >= 3 AND group3 >= 1) → upgrade one level
  • mediumUpgrade: group2 >= 2 OR (group1 >= 1 AND group2 >= 1) → upgrade one level
  • no signal: keep rule grade unchanged; report "词表命中不足" with low confidence
• confidence = min(1, (group1Hits + group2Hits + group3Hits + searchEvidence) / 6)

5. Semantic source tag:

• standards — grading based on term hits only
• standards_with_semantic_search — grading also used semantic_search evidence

6. Final grade synthesis (synthesizeFinalGrade):

• If rule grade = semantic grade → final = rule grade, no human review needed
• If gap >= 2 levels → final = rule grade, needs_human_review=true, reason: semantic_major_divergence or semantic_extreme_divergence
• If rule=C and semantic=D → final = rule grade, needs_human_review=true, reason: semantic_downgrade_review
• If gap = 1 level → final = semantic grade, no human review needed
• Upgrade (B→A, C→B) and downgrade (B→C) are auto-adopted without entering human review sheet

7. Output fields per item:

• rule_grade: letter from rule-based classifier
• semantic_grade: letter from this skill
• semantic_reason: Chinese explanation string
• semantic_confidence: 0.00–1.00
• semantic_source: standards or standards_with_semantic_search
• final_grade: letter from synthesizeFinalGrade
• needs_human_review: boolean
• disagreement_type: string (empty if no disagreement)

8. Integration with med-daily-triage:

• 三列输出 (规则等级, 语义等级, 最终等级) go into 隔日报.xlsx "每日反馈" sheet
• needs_human_review=true items go into "需人工复核" sheet

9. Guardrails:

• Never downgrade an item solely because term hits are zero — insufficient evidence keeps rule grade
• Semantic evidence is supplementary; it cannot override explicit user feedback (see med-query-learning)
• All three term groups come from config; skill must not hardcode domain vocabulary