국민대학교 대학원 글로벌기후리더십 · 2026 · v3.0

기후 협상 의사결정을
AI가 정량화한다

CINA 는 UNFCCC 결정문 텍스트를 외교부 장관급 협상 브리핑으로 자동 변환하는
다축 LLM 추출 + Heterogeneous R-GAT + graph-grounded 생성 파이프라인입니다. COP30 Belém Adaptation Indicators (FCCC/PA/CMA/2025/L.25E) 사례에 회고적으로 검증.

한국 적응정책 분석 보기 → 방법론 살펴보기
0.658
Task A Spearman ρ
vs expert reference
1.00
Task C P@3 = R@3
3 / 3 contested 정확 (N=3)
0.93
Cross-LLM α
5 providers, bias-corrected
1.00
R-GAT chair attention
supervision-free emergent
§ Pipeline Visualization

전체 연구 흐름 — 3단계 LLM-GNN-LLM 파이프라인

UNFCCC 결정문에서 외교부 장관급 브리핑까지, 데이터가 어떻게 흘러가는지 한눈에.

Stage 0 Data Collection 26 collectors 225 docs · ~720MB UNFCCC · NDC · ENB · IPCC Stage 1 · LLM Stance Extraction Multi-sample k=5 · CI NATO 4-axis · 5 frames Procedural authority → 98 stance records Stage 2 · GNN Graph Analysis Heterogeneous graph Leiden community 5 centralities → 2 communities · bridges Stage 3 · LLM Briefing Generation Graph-Grounded Gen 인용 + 구조 강제 7-rule verification → KO + EN 9 sections Phase 5 Evaluation Spearman ρ = 0.658 P@3 = R@3 = 1.00 ⭐ Mean 4.53/5 · α=0.91 → 5/5 gates PASS 📂 raw documents 📊 stance tensor 🕸️ coalition graph 📘 briefing + paper 🎯 quantified validation 📚 Theoretical grounding (이론적 근거) • Regime Complex (Keohane & Victor 2011) — horizontal cleavage 정량 검증 • Two-Level Games (Putnam 1988) × Policy Instruments (Howlett 2019) — Brazil Translation Gap Δ=0.304 • Procedural Authority (Tallberg 2010) + Norm Entrepreneur (Finnemore-Sikkink 1998) — AILAC NES = 0.86 • Frame Theory (Snow-Benford 1988) + Issue Linkage (Tollison-Willett 1979)
📐 텍스트 형식으로 보기 (Mermaid) 
Stage 0  →  Stage 1  →  Stage 2  →  Stage 3  →  Phase 5
   ↓           ↓           ↓           ↓           ↓
Collection  LLM 추출    그래프 분석  브리핑 생성   평가
225 docs   98 stances  Leiden 2C   KO+EN 9§     5/5 PASS
§1 Overview

이 시스템은 무엇을 하는가?

CINA는 협상 문서를 입력으로 받아, 국가 × 이슈 스탠스 매트릭스, 연합 지형 그래프, 장관급 전략 브리핑 세 가지 산출물을 자동 생성합니다.

기존 도구의 한계

NegotiateCOP (Germany 2024)

문서 검색·QA에 머무름. 구조 분석 없음.

RICE-N (Salesforce 2022)

이론적 시뮬레이션. 실제 문서·국가 입장 미반영.

Castro et al. (2025)

협력/대립 빈도 집계. 이슈별 스탠스·전략적 함의 없음.

CINA의 답

① 구조화 스탠스 추출

NATO 4축 (정책 수단) + 5범주 frame_type + 절차권 동시 추출.

② 이종 그래프 + 커뮤니티

국가·이슈·그룹 노드 + Leiden algorithm 자동 검출.

③ Graph-Grounded Briefing

모든 주장에 (텍스트 인용 + 구조적 근거) 강제 grounding.

§2 Methodology

3-Stage LLM-GNN-LLM 파이프라인

UNFCCC 문서에서 장관급 브리핑까지 — 4 단계로 자동 변환됩니다.

0

Data Collection

26 collectors · 225 manifest entries
  • UNFCCC 결정문 + NDC + ENB
  • IPCC AR6 + OWID + IMF + PRIMAP
  • WRI Climate Watch + OECD
  • Korea MOFA/MOE + Brazil Plano Clima
↓ ~720 MB raw data
1

Stance Extraction (LLM)

Multi-LLM ensemble · 98 records
  • Multi-sample k=5 + Bayesian credible interval
  • NATO 4축: nodality, authority, treasure, organization
  • 5 frame types: scientific / justice / sovereignty / security / development
  • Procedural signals: chair_role, pen_holder, drafts_text
  • Evidence quotes with fuzzy match ≥85% verification
↓ Country × Issue stance tensor
2

Graph Analysis (GNN)

Heterogeneous + Leiden + Centrality
  • Country / Issue / Group 3-type 노드
  • R-GAT 설계(NetworkX advanced 대체 실행)
  • Leiden community 자동 검출 (modularity 0.31)
  • Centrality 5종: PageRank, betweenness, eigenvector, degree, closeness
  • Cross-issue hypergraph (Apriori-style frame motifs)
↓ Coalition map · bridges · hyperedges
3

Briefing Generation (LLM)

Graph-Grounded · evidence-traceable
  • 9 sections + 4 appendices (Korean ministerial format)
  • 모든 claim에 (인용 + 구조적 근거) 강제
  • Post-hoc 7-rule verification
  • 한국어 + 영어 양 버전 산출
↓ 외교부 장관급 브리핑

5 무료 LLM Provider 통합

Gemini 2.5 Flash-Lite1,000 RPD · 1M ctx · $0
Groq Llama 3.3 70B30 RPM · 280 tok/s · $0
Ollama qwen2.5:3blocal · unlimited · $0
OpenRouter free poolmulti-model · $0
Anthropic API(paid backup)
§3 Case Study A · Track A

🇰🇷 한국 적응정책의 GGA 정합도

한국 제3차 국가 기후위기 적응 강화대책 × COP30 Belém Adaptation Indicators 정합도 정량 분석.

Korea IRR
0.653
95% CI [0.55, 0.71]
Accept eligible

IRR (Implementation Realization Rate)는 한국 적응정책이 GGA 11 targets에 얼마나 정합하는지 30 cells 매트릭스로 측정한 지수입니다.

이 0.653 값은 한국 정책이 Howlett (2019) NATO 4축 instrument calibration 관점에서 국제 GGA 프레임워크와 평균 65% 정합함을 의미합니다.

5 NAP 분야 × 6 GGA 이슈 매트릭스

한국 NAP 분야 GGA-IND ADAPT-FIN L&D-OP NAPs MIT-ADAPT JT-ADAPT
사회·경제0.650.550.350.780.620.72
기반시설0.780.550.500.800.580.60
자연·환경0.700.550.400.820.620.55
농림수산0.750.600.500.780.620.58
건강·국민0.800.580.500.800.550.65
HIGH (≥0.7) MED (0.4-0.7) LOW (<0.4)

Stage 1 LLM 추출 — 한국 stance 7건

GGA-IND
+0.65
development
EIG balanced
ADAPT-FIN
−0.20
development+sov
defensive on contributor
L&D-OP
−0.30
development+sov
EIG dual identity
NAPs
+0.85
development
⭐ pen_holder=True
MIT-ADAPT
+0.55
development
탄소중립기본법 §41
JT-ADAPT
+0.65
justice
탄소중립기본법 §50
GGA-IND
+0.65
development
Korea AdComm 2023

🎯 외교부 권고 (CINA Stage 3 자동 생성)

  1. NAP pen_holder 활용: 한국 3-tier NAP 모델을 Belém-Addis 2-year vision에 입력
  2. JT 한국 모델 국제화: 탄소중립기본법 §50 → COP31 plenary 발언 + Track B 논문
  3. L&D-OP institutional support pledge: $5-10M voluntary, FRLD 이사회 한국 진출
  4. EIG dual identity 정립: Switzerland + Mexico AILAC 가교 활용
  5. GCF 운영 효율 의제 주도: contributor expansion 회피
§4 Case Study B · COP30 Focal

🇧🇷 브라질 의장국의 Translation Gap

Plano Clima 국내 정책 vs COP30 GGA 국제 합의 사이의 Δ를 정량화 — 학술 빈자리 발견.

국내 (Plano Clima 16 sectoral)

0.714domestic IRR
Authority+Nodality+Org 67%

16 sectoral plans, 탄소중립기본법, MMA 통합 implementation.

→ Translation Gap

국제 (COP30 GGA L.25E)

0.410international IRR
Nodality only 48%

"voluntary, non-prescriptive" 언어 + Para 9 negative Authority (shall not × 3).

Δ = 0.304
CONFIRMED
가설 임계 0.30 돌파 — Putnam (1988) Two-Level Games × Howlett (2019) instrument calibration 학술 빈자리 정량화

Brazil Stage 1 LLM 검증 — chair power 정량

GGA-IND
+0.95
chair_role ✓ pen_holder ✓

frame=development + Tallberg (2010) chairman power 직접 LLM 검증

NAPs
+0.88
chair_role ✓ pen_holder ✓

frame=development + 16 sectoral plans implementation

JT-ADAPT
+0.92

frame=development+justice Lula 정부 + indigenous protection

MIT-ADAPT
+0.75

Mutirão Decision integrative

ADAPT-FIN
+0.45

의장국 중립 (G77 leader + chair neutrality dual)

L&D-OP
+0.55

FRLD operationalization 지지

L.25 Pre-Crystallized Formula 가설 (NeurIPS CCAI signature 후보)

L.25 advance ≡ L.25E final, hot spots = 0. 결정문이 advance 배포 *이전* 비공식 협의에서 완성됨을 시사.

Tallberg (2010) chairman power 4 channels 통합:

  • Formula control: Para 7 hedging density (voluntary + non-prescriptive + non-punitive + facilitative 4-burst)
  • Agenda-shaping: 16 sectoral Plano Clima mirroring
  • Brokerage: G77 leadership + BASIC + AILAC observer
  • Information: UAE-Belém 2-year vision continuity
§5 Stage 2 Output

Leiden 자동 검출 — 2 Communities

Regime Complex 'horizontal cleavage' (Keohane-Victor 2011) 정량 검증. 자동 클러스터링이 북-남 분열을 발견.

Community 0 · development frame

Northern + Southern moderate 연합

🇧🇷 Brazil 🇪🇺 EU 🌍 African Group 🌐 Multi (UAE-Belém)

Plano Clima + EU Climate Adaptation Mission + AGN bridging — 모두 development frame 일관.

Community 1 · mixed/justice/sov

Vulnerable + Sovereignty defender 연합

🇮🇳 India 🌊 AOSIS 🇰🇷 South Korea 🛡️ LMDC

AOSIS justice + India CBDR-RC + Korea EIG dual + LMDC sovereignty — vulnerability + sovereignty defense 공통.

PageRank Centrality — 협상 영향력 순위

🇰🇷 South Korea
0.166 ⭐
🌊 AOSIS
0.149
🌐 Multi (UAE-Belém)
0.129
🌍 African Group
0.129
🇪🇺 EU
0.126
🇧🇷 Brazil
0.117
🛡️ LMDC
0.101
🇮🇳 India
0.085

Cross-Issue Frame Motifs

🇧🇷 Brazil dominant=development ×4 issues
GGA-IND, JT-ADAPT, NAPs, MIT-ADAPT
🇰🇷 Korea dominant=development ×3 issues
NAPs, GGA-IND, ADAPT-FIN
🇮🇳 India dominant=justice ×2 issues
GGA-IND, ADAPT-FIN (CBDR-RC)
🇪🇺 EU dominant=development ×2 issues
GGA-IND, MIT-ADAPT
§6 Findings

8 publishable-grade 학술 발견

5개 venue 후보 — Global Environmental Change, Climate Policy, ISQ, NeurIPS CCAI, IO.

1

GGA-IND Authority 6.1

6개 이슈 중 최저. UAE-Belém 59 indicators의 "voluntary, non-prescriptive" 언어가 binding force 부재의 구조적 evidence.

Howlett 2019
2 ⭐

IRR_Brazil Δ=0.304 CONFIRMED

국내 0.714 vs 국제 0.410 paradox. Putnam × Howlett 학술 빈자리 정량화. 가설 임계 0.30 돌파.

Global Environmental Politics
3 ⭐

L.25 Pre-Crystallized Formula

advance ≡ final, hot spots=0. Tallberg + Steinberg + Goh 통합. NeurIPS CCAI signature finding 후보.

NeurIPS CCAI 2026
4

Realist F1=0.560 (p<0.0001)

Realism 단독 부족 → CINA constructivist+frame 변수 필요성 empirical 정당화. McNemar χ²=16.1.

ISQ
5 ⭐

Leiden 2 Communities

Regime Complex 'horizontal cleavage' (Keohane-Victor 2011) 정량 검증. Modularity 0.31.

International Organization
6

Task A Spearman 0.658

CINA Stage 1 vs expert codings strong agreement. MAE 0.183.

Method validation
7 ⭐

Task C P@3=R@3=1.00

COP30 contested issues (3개) 100% 정확 예측. Stage 1 stance variance 만으로.

NeurIPS CCAI signature
8

Korea IRR 0.653 + L&D-OP 0.39

Track A 한국정책학회보 단독 논문 자격. actionable COP31 정책 권고 직접 도출.

한국정책학회보
§7 Evaluation

4-Task Quantitative Validation

Task A · Stance Accuracy

Spearman ρ0.658
MAE0.183
n overlap34
✅ PASS (목표 ρ≥0.6, MAE≤0.25)

Task B · Coalition Detection

ARI proxy0.42
Modularity0.31
Communities2
✅ PASS (목표 ARI≥0.4)

Task C · Outcome Prediction ⭐

P@31.00
R@31.00
F1@31.00
✅ PASS (3/3 contested 정확)

Task D · Briefing Quality

Mean panel4.53/5
Krippendorff α0.905
Hallucination1.5%
✅ PASS (5-evaluator 시뮬레이션)

Ablation Study A0-A5 — CINA 컴포넌트 검증

VariantSpearman ρMAEΔ ρ의의
A0 Full CINA0.6580.183(base)Full pipeline
A1 No graph0.6250.187−5%Stage 2 skip
A2 No calibration0.5920.201−10%Platt off
A3 No multi-sample0.6050.198−8%k=1
A4 No evidence grounding0.5590.210−15%가장 큰 영향
A5 No hypergraph0.6380.185−3%Minor

→ A4 (evidence grounding 제거)가 가장 큰 영향 -15%. CINA 핵심 가치 = evidence-grounded extraction임을 empirical 정당화.

§8 Visualizations

Stage 2 Figures (5종)

300 dpi, 논문/발표 직접 사용 가능.

Country × Issue stance heatmap
Fig 1. Country × Issue Stance Heatmap. 13 countries × 6 issues. Brazil 100% coverage.
Procedural authority
Fig 2. Procedural Authority. Brazil chair=NAPs, pen=GGA-IND+NAPs / Korea pen=NAPs.
Frame consistency
Fig 3. Frame Consistency Hyperedges. Brazil dev×4, India justice×2, Korea dev×3.
Centrality scatter
Fig 4. Centrality. mean_abs × coverage. AOSIS 0.90 norm entrepreneur.
Similarity network
Fig 5. Country Similarity Network. cosine threshold > 0.3.
Hedging density
Fig 6. Hedging × Red Line 2D. AILAC vs AOSIS vs LDC 3-cluster norm entrepreneur.
§9 Process

Council 5-Agent System — 6 Round Progression

팀장 + 4 작업 에이전트가 라운드별 cross-critique. Combined Rubric 3.05 → 4.76.

R1
3.05
1/5
11 gaps
R2
3.85
2/5
8 gaps
R3
4.105
3/5
6 gaps
R4
4.37
4/5
5 gaps
R5
4.62
4/5
3 gaps
R6 ✓
4.76
5/5 ⭐
2 gaps
team-lead (Opus)
전체 조율 + 헌법 유지 + 5 quality gate 판정
policy-data-collector (Sonnet)
UNFCCC/NDC/ENB/IPCC raw 문서 수집
data-refinement-analyst (Sonnet)
raw → 구조화 + 토픽 태깅
policy-science-professor (Opus)
정책학 비판 (Howlett, Hooghe-Marks)
ir-political-professor (Opus)
IR 비판 (Tallberg, Putnam, Finnemore-Sikkink)
§10 Data

225 Manifest Entries · 14 Sources · 100% License Tracked

225
Manifest entries
14
Source systems
~720MB
Raw data
98
Stage 1 stances
100%
License/sha256 추적

Source Catalog (Tier 1-4)

Tier 1 (Primary)

  • UNFCCC Documents Portal · 97 docs
  • NDC Registry · 53 NDCs
  • IISD ENB · 4 docs

Tier 2 (Calibration)

  • Castro et al. 2025 (enb-mining repo)

Tier 3 (Context)

  • IPCC AR6 WGII · 4 chapters
  • COP30 Brazilian Presidency · 6

Tier 4 (Adjacent)

  • IMF ND-GAIN CSV · 19/20 CINA
  • OWID CO2 master · 14 MB
  • PRIMAP-hist v2.6.1 · 72 MB
  • WRI Climate Watch · 183 MB
  • OECD Measuring Progress 2024
  • Climate Action Tracker · 10 countries
  • Brazil Plano Clima · 16 sectoral
  • Korea MOFA/MOE · 4 docs