CINA 프로젝트의 전 산출물 — 학술 페이퍼, 장관급 브리핑, 정량 평가 보고서, Stage 2 시각화, 데이터셋. 모든 항목은 GitHub repository에서 직접 열람 가능합니다.
Track A (수업 제출 — 장관급 브리핑) + Track B (학술 투고 — 페이퍼) 듀얼 트랙 산출물 일체.
3,428 words · 23 references · 7 sections + KO summary. Target venue: Global Environmental Change / NeurIPS CCAI 2026 / 한국정책학회보. Inline figure refs (1-6).
외교부 장관급 협상 전략 브리핑 (한국어). 9 섹션 + 4 부록(증거 추적성, 데이터 lineage, 불확실성, LLM provider attribution).
English version of the ministerial briefing — 310 lines. Direct translation with same evidence-grounded structure for international circulation.
IRR_Korea = 0.653, CI [0.55, 0.71]. 30/30 cells (5 NAP 섹터 × 6 GGA 이슈) Korean NAP × GGA crosswalk 적용.
Brazil Translation Gap Δ = 0.304 CONFIRMED (Plano Clima 67% → GGA voluntary 48%). Putnam × Howlett gap 정량화 — NeurIPS CCAI signature finding.
L.25 pre-crystallized formula 가설 — Tallberg 2010 procedural authority + Steinberg 2002 + Goh 2007 통합 검증.
Realist B0 F1 = 0.560, McNemar p<0.0001, κ=0.216. Constructivist + frame variables 경험적 정당성 확보.
AILAC norm entrepreneur 가설 — Finnemore-Sikkink (1998) 4 criteria 3.5/4 PASS. NES = 0.86. Tactical diplomatic framing 정량화.
Phase 5 4-task 정량 검증. Spearman ρ=0.658, P@3=R@3=1.00, Mean 4.53/5, α=0.905. Ablation A0-A5 + 5/5 PASS.
Korean National Adaptation Plan × GGA 5 NAP × 6 GGA = 30 cells. IRR_Korea 산출 기반.
Brazil 선정 근거 + Adaptation 섹터 정당성 + COP30 6-issue 매트릭스 (수업 과제 사전 분석).
모든 figure는 data/processed/figures/ 또는 deliverables/에서 생성 — matplotlib + networkx, reproducible via src/stage2_graph/generate_figures.py.
13 countries × 6 issues stance tensor (Stage 1 산출). 적색=opposing, 청색=supporting. 페이퍼 §3 anchor figure.
Tallberg 2010 framework: chair_role + pen_holder + drafts_text 신호 분포. Brazil COP30 chair authority 정량화.
5-frame typology (scientific/justice/sovereignty/security/development) × 그룹별 빈도. AOSIS=justice, Saudi+India=sovereignty.
5 centrality (PageRank·betweenness·eigenvector·degree·closeness) Top-K. Brazil 의장국 + EU 펜홀더 우위.
Leiden 2 communities: C0={Brazil, Multi, AGN, EU} development / C1={AOSIS, India, Korea, LMDC} mixed. Keohane-Victor 'horizontal cleavage' 정량 검증.
AILAC (high hedging + mod RL — tactical), AOSIS (low hedging + high RL — principled), LDC (mod + low — vulnerable). Finnemore-Sikkink norm typology 시각 검증.
추가 그룹 라벨링 + 95% confidence ellipse 적용.
| Ablation | Removed | Spearman ρ Δ |
|---|---|---|
| A0 (full) | — (baseline) | 0.658 |
| A1 | Bayesian CI | −0.05 |
| A2 | NATO 4-axis | −0.07 |
| A3 | Frame typology | −0.04 |
| A4 | Evidence grounding | −0.15 |
| A5 | Multi-LLM ensemble | −0.06 |
→ Evidence grounding (A4) is the single most load-bearing component — 페이퍼 §5.3 핵심 발견.
raw 평가 데이터 (machine-readable). Task A/B/C/D + Ablation 전체 메트릭 + per-record breakdowns.
Phase 5 4-task 평가 재현 스크립트. python -m src.evaluation.run_full_evaluation으로 모든 메트릭 재생성.
논문 §2-3 (이론·방법론)의 직접 구성 요소. docs/ 폴더 내 15개 markdown 일체.
Regime Complex (Keohane-Victor) · Two-Level Games (Putnam) · Issue Linkage (Tollison-Willett) · Epistemic Communities (Haas)
View →Three-stage pipeline 공식 정의. Stage 1 LLM extraction · Stage 2 graph analysis · Stage 3 generation.
View →Multi-sample k=5 + Bayesian credible interval + Platt calibration + evidence_quote 강제.
View →R-GAT 이론 사양 + Leiden community + Apriori-style hyperedge mining.
View →Graph-Grounded Generation + 7-rule post-hoc verification.
View →chair_status + drafts_text edge type + chair authority weighting.
View →225 raw 문서 metadata 추적 (license, sha256, source, version, collected_at).
View JSONL →웹사이트 시각화에 임베딩된 실데이터. Stage 1 stances + Leiden + IRR + Task A-D.
View JSON →전 파이프라인은 MIT 라이선스로 GitHub에 공개 — 재현 가능한 reproducibility 우선.
End-to-end CLI: python -m src.pipeline --country Brazil --cop 30 --sector adaptation --stages 1,2,3.
5 LLM provider abstraction (Gemini/Groq/Ollama/OpenRouter/Anthropic) + multi-sample + calibration.
Folder →NetworkX heterogeneous graph + Leiden (igraph) + 5 centralities + figure 생성.
Folder →5-agent council Phase A-E orchestration + state_manager + quality gate.
Folder →26 collectors (UNFCCC, NDC, ENB, IPCC, IMF, PRIMAP, WRI, Brazil, Korea, Castro 2025).
Folder →