antinet-four-color-cards

Four-Color Card Generation Skill (Tongzhengsi + Jianchayuan + Chengxiangfu)

How to Use

• Mainly invoked by the Tongzhengsi Worker to generate blue cards (facts) / green cards (explanations).
• Yellow cards (risks) are handed to the Jianchayuan for review of overclaiming; red cards (actions) are handed to the Chengxiangfu to complete strategic recommendations.
• The four cards are uniformly recorded by the Taishige to ensure each card is traceable.

Input

• markdown: structured document produced by Mijuanfang parsing
• intent: user intent / query objective
• llm_available: whether the LLM is available (affects the yellow card generation path)

Output

• blue_card: Fact card — a traceable statement, with source snippets attached
• green_card: Explanation card — mechanism/background explanation of the facts
• yellow_card: Risk card — overclaiming, uncertainty, and potential bias (reviewed by Jianchayuan)
• red_card: Action card — next-step recommendations (completed by Chengxiangfu)

Dependencies

• LLM API (generates blue/green/red cards)
• Overclaiming detection rule set (provided by Jianchayuan, used for yellow cards)
• Parsing confidence (from doc-parse; low confidence triggers a manual review flag)

Failure Handling

• LLM unavailable → degrade to the rule engine to generate cards, and explicitly mark yellow/red cards as "LLM not enabled".
• Jianchayuan review timeout → yellow cards are temporarily delivered using rule engine results, and mark pending_review。
• Missing source → the corresponding fact is marked "no source" in the blue card, and no assertion is generated directly.

Reuse Value

• Methodology is general: scientific research analysis, financial research reports, and legal document analysis can all apply the four-color card framework.
• Strong auditability: each card carries its source and generation method, naturally fitting the observability/audit review dimensions of Agent Infra.

Semifinal Code Package Execution (runnable package)

• Real entry point:scripts/run_four_color_cards.py --stage {extract|review|propose}
• Three subtask mappings:
  • extract → Tongzhengsi comm.tongzhengsi.TongZhengSiAgent（blue card: fact extraction）
  • review → Jianchayuan audit.jichayuan.JianChaYuanAgent（green card: Gap/explanation）
  • propose → Chengxiangfu strategy.chengxiangfu.ChengXiangFuAgent（red card: action recommendations）
• Run example:
  • python skills/four-color-cards/scripts/run_four_color_cards.py --stage extract
  • python skills/four-color-cards/scripts/run_four_color_cards.py --stage review
  • python skills/four-color-cards/scripts/run_four_color_cards.py --stage propose
• Artifacts:examples/snse_survey/skill_outputs/four_color_<stage>.json
• Iron rule of traceability: green cards cite blue cards, red cards cite green/blue cards, and nothing may be stored without a source (enforced by card_model).