med-daily-triage

name: med-daily-triage description: 每日分级与反馈表生成技能。Use when classifying merged papers into A/B/C/D and writing 隔日报.xlsx with clickable dropdown fields and Chinese-first review text.

med-daily-triage

1. Classify merged candidates using rule-based grading:

• A课题相关: directly relevant to current core project question.
• B专题相关: clearly relevant to subtopic or adjacent topic.
• C领域相关: broader field relevance, lower short-term priority.
• D无关: low relevance; audit only, never written to Zotero.

2. After rule-based grading, apply semantic grading (see med-semantic-grading) for second-pass review.

3. For database-sourced items, apply the journal quality gate backed by easyscholar partition/rank information before final writeback. This extra gate is advisory but auditable, and it should only affect database retrieval items, not RSS items.

4. Generate 隔日报.xlsx with two sheets that do NOT overlap:

Sheet 1: "每日反馈"

• Exclude all D无关 items AND all items where needs_human_review=true
• Required columns: 英文标题, 标题翻译, 规则等级, 语义等级, 最终等级, 期刊/来源, 反馈, 评价
• Three grade columns (规则等级, 语义等级, 最终等级) output only letters A/B/C/D, no explanatory text
• Auto-adopted adjustments (B→A, C→B, B→C) appear here via 规则等级/语义等级/最终等级 but do NOT enter 人工复核

Sheet 2: "需人工复核"

• Only items where needs_human_review=true (e.g. C→D blocked by policy)
• Additional column: 人工确认等级 for user to fill (A/B/C/D dropdown)
• 规则等级, 语义等级, 最终等级 are read-only evidence

Invariant: Sheet1 rows + Sheet2 rows = all ABC items actually written to Zotero that day.

5. Data source rule:

• desktop_daily_review_source.json must contain only items from Stage2 mcp_writeback_summary.writeback_items
• Stage1 full ABC is the candidate pool, not the 隔日报 data source
• If Stage2 fails/skips, desktop source keeps original full set

6. Clean journal/source display:

• Remove suffixes: : Latest Articles (...), : Latest Articles, : Table of Contents, - Wiley Online Library, - Wiley, ScienceDirect Publication:
• Remove noise values: Latest Results, Wiley, Wiley Online Library, ScienceDirect, ACS Publications

7. Apply dropdown validation:

• Columns C/D/E (规则等级/语义等级/最终等级): A,B,C,D
• Column G (反馈): keep,drop,upgrade,downgrade
• Column K in "需人工复核" sheet (人工确认等级): A,B,C,D

8. Column mapping notes:

• 反馈 column = renamed from feedback
• 评价 column = renamed from comment
• Removed columns: 来源等级, 已处理时间, 处理状态, 备注

9. Guardrails:

• Never fabricate dose/time/results/stats.
• If confidence is low, mark review queue instead of asserting evidence.
• Keep stable traceability for A/B items for star migration scanning.
• Integrate feedback-driven tightening (see med-query-learning).
• Chinese title must be machine-translated; if translation fails, keep English and log failures in run report.
• Feedback strength is not uniform: keep is a no-op for placement, upgrade is a positive reinforcement signal, downgrade is a weaker negative signal, and drop is a stronger negative signal.