| title | From Text to Strategy — A Multi-LLM Climate Issue-Network Analysis (CINA) Pipeline, Retrospectively Validated on COP30 Adaptation Outcomes | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| author |
|
|||||||||||
| generated_at | 2026-05-03 | |||||||||||
| target_venue |
|
|||||||||||
| status | Polished draft — citations integrated, figures inline-referenced (not yet submitted) | |||||||||||
| license | CC BY 4.0 (paper) / MIT (code at github.com/zxsa0716/cina) | |||||||||||
| language | EN + KO summary |
From Text to Strategy — A Multi-LLM CINA Pipeline, Retrospectively Validated on COP30 Adaptation Outcomes
Climate negotiation analytics have so far split into three complementary but disconnected approaches: (a) retrieval-augmented document interfaces (e.g. NegotiateCOP, Germany 2025), which surface and compare individual country positions but do not jointly model policy-instrument calibration or coalition structure; (b) game-theoretic simulators (RICE-N, Zhang et al. 2025), which generate strategic dynamics without grounding in negotiation texts; and (c) interaction-frequency datasets (Castro et al. 2025, Nature Scientific Data; Castro et al. 2025, Environmental Sociology), which aggregate ENB-coded cooperation/conflict counts and apply dynamic network analysis to recover alliances. We introduce CINA (Climate Issue-Network Analysis), a methodology that complements these approaches by combining four signals into a single LLM-based stance extraction step: (i) stance scores with Bayesian credible intervals from k=5 ensemble sampling, (ii) NATO 4-axis policy instruments (Hood 1983; Howlett 2019; Capano et al. 2025), (iii) frame types (5 categories; Snow & Benford 1988), and (iv) procedural-authority signals (chair_role, pen_holder, drafts_text; Tallberg 2010). Downstream, a heterogeneous country × issue × group graph is analysed in two ways: a NetworkX + Leiden realisation (Traag, Waltman & van Eck 2019) and a multi-task heterogeneous R-GAT (PyTorch; Schlichtkrull et al. 2018; Veličković et al. 2018) trained jointly on stance regression, coalition classification, and contested-issue prediction. We compute a cross-LLM Krippendorff α = 0.876 (raw) / 0.933 (bias-corrected, after per-LLM mean-centering) across five providers, providing a reliability bound that is partially decoupled from shared-model bias. A graph-grounded generation discipline forces every output claim to cite both a verbatim evidence quote and a structural fact; ablating this reduces Spearman ρ by 15 percentage points, the largest single component effect. We pilot CINA on the COP30 Belém Adaptation Indicators outcome (FCCC/PA/CMA/2025/L.25E, November 2025) as a single retrospective case. Across a 4-task evaluation on n=98 stance records spanning 13 countries × 6 issues, CINA reaches Spearman ρ = 0.66 against expert reference codings (Task A); the R-GAT extension reaches validation Spearman ρ = 0.71 with coalition accuracy 0.77 and contested P@3 = 0.67, and crucially its learned attention weights — without explicit supervision on procedural authority — are dominated by chair edges (mean attention 1.00) over similarity edges (0.28), an emergent recovery of Tallberg's (2010) chair-channel hypothesis. CINA correctly identifies 3 of 3 COP30 contested issues from stance variance alone (Task C; N = 3). We explicitly note that under a hypergeometric chance baseline (selecting 3 of 6 issues at random when 3 of 6 are contested), the probability of 3/3 correct prediction by chance is C(3,3)·C(3,0)/C(6,3) = 1/20 = 0.05; the observed result is therefore 20× the chance baseline but the small N permits only a single-case "encouraging signal" claim, not generalisable accuracy. Prospective validation on COP31 is required. A simulated 5-persona expert panel scored generated briefings at 4.53/5 mean (Task D; we flag this as a methodological limitation pending real-expert validation). As an empirical observation we report a Brazil domestic-international policy-instrument divergence of Δ = 0.304 between Plano Clima (NATO 4-axis 67 %) and the L.25E international text (Nodality-only 48 %), proposed as a novel single-case quantification at the intersection of Two-Level Games (Putnam 1988) and instrument-calibration theory (Howlett 2019); a difference-in-differences / synthetic-control / IV identification strategy is pre-specified to test causal generalisation. The four hypotheses for COP31 prospective validation are pre-registered with code freeze planned for 2026-09-01. CINA is released under MIT/CC BY 4.0 at github.com/zxsa0716/cina.
Keywords: climate negotiations, large language models, graph neural networks, regime complex theory, two-level games, instrument calibration, COP30, adaptation, Korea, Brazil
The 2025–2026 climate adaptation negotiation cycle generated nearly 300 official UNFCCC documents, hundreds of national submissions, and the first formal Belém Adaptation Indicators package — 59 voluntary indicators across seven thematic targets, accompanied by a decision to triple adaptation finance to USD 120 billion by 2035 (FCCC/PA/CMA/2025/L.25E; FCCC/PA/CMA/2025/L.24). The volume and complexity of this textual record have motivated several recent computational approaches, each strong in one respect but limited in another. NegotiateCOP (GIZ, 2025) is a public retrieval-augmented interface that lets users search submissions and compare positions side by side, but does not jointly model policy-instrument calibration, frame typology, or coalition structure. Castro et al. (2025) released a Nature Scientific Data dataset of cooperation/conflict counts coded from Earth Negotiations Bulletin (ENB) reports across COP1–COP29, and a companion Environmental Sociology paper (Castro et al. 2025) applied dynamic network analysis to that dataset; that line of work captures interaction frequencies but not the policy-instrument or framing signals embedded in the underlying texts. RICE-N (Zhang et al. 2025) simulates negotiation dynamics with multi-agent reinforcement learning, producing strategic insights without document grounding.
This paper proposes a complementary methodology, CINA (Climate Issue-Network Analysis), that combines four signals into a single LLM-based stance extraction step — Bayesian-uncertainty-quantified stance scores, NATO 4-axis policy-instrument calibration (Hood 1983; Howlett 2019; Capano et al. 2025), 5-frame typology (Snow & Benford 1988), and procedural-authority signals (Tallberg 2010) — and feeds the resulting stance tensor into a Leiden-based heterogeneous graph analysis (Stage 2), and from there into a graph-grounded generation step that produces ministerial-grade briefings under verifiable evidence-and-structure dual-grounding (Stage 3). The contribution is methodological: each individual signal has antecedents in the literature, but their joint extraction within a single multi-axis schema, validated against an actual COP outcome (the L.25E text adopted in November 2025), is to our knowledge novel within the climate-negotiations domain.
We frame three research questions:
- RQ1 (Methodology): Can a single LLM extraction step, applied at the (country, issue) granularity, recover stance direction together with policy-instrument calibration, frame typology, and procedural-authority signals at acceptable accuracy?
- RQ2 (Empirical observation): Does Leiden community detection on the resulting stance tensor produce a partition consistent with the regime-complex horizontal-cleavage hypothesis (Keohane & Victor 2011)?
- RQ3 (Single-case retrospective validation): When CINA is given the pre-COP30 corpus only, does it correctly identify the three issues that turned out contested at COP30, and does it surface a measurable domestic-international policy-instrument divergence in the chair country (Brazil)?
We position the present paper as a preliminary methodology paper with single-case retrospective validation. Sections 2–3 review the literature and methodology. Section 4 reports empirical observations. Section 5 reports the 4-task evaluation, including explicit limitations of a simulated-panel briefing-quality task (Task D) and a small contested-issue sample (n = 3). Section 6 discusses Korean policy implications, and Section 7 concludes with future work — in particular, expansion to n = 300+ stance records, real-expert validation, and PyTorch-based graph-attention learning.
CINA's design grounds in four complementary IR / policy-analysis traditions, mapped one-to-one onto pipeline modules.
Climate governance is not a single regime but a loosely-coupled regime complex: UNFCCC, Paris Agreement, Loss-and-Damage Fund, Green Climate Fund, IPCC, and bilateral arrangements coexist with overlapping but non-identical rule sets. Implication: a country's "stance" on adaptation is not a scalar — it is a vector across issues, each issue exposing a different rule subset. Our Stage 1 extraction is per-(country, issue), not per-country.
International negotiators bargain simultaneously at the international table (Level I) and the domestic ratification arena (Level II). Agreement requires the intersection of both win-sets. Implication: a country's submission encodes flexibility signals (where it can move) and red lines (where domestic politics fixes the boundary). Our schema captures both as separate fields.
Single-issue impasses can be unblocked by bundling issues so that distinct preference orderings yield Pareto improvements. Implication: we mine cross-issue hyperedges with Apriori-style frame motifs (Section 3.2).
Technically complex regimes admit epistemic communities — networks of credentialed experts whose causal beliefs guide policy. Implication: we measure expert-proposal vs political-text divergence as an epistemic_divergence score, and we treat the COP30 Belém "Rube Goldberg" gap (where the chair-drafted final text diverges from the Technical Expert Group recommendation) as the canonical retrospective validation case.
Tallberg's four chairmanship channels — formula control, agenda-shaping, brokerage, information — are operationalised as Boolean fields (is_chair_role, is_pen_holder, drafts_text_for_issue) that the LLM extracts directly from L-document headers. Steinberg's "consensus-shaping" and Goh's "informal pre-cooking" provide the conceptual frame for our pre-crystallized formula hypothesis (Section 4.3).
| Tier | Source | Count | License |
|---|---|---|---|
| Tier 1 | UNFCCC Documents Portal | 97 | UN Open License |
| Tier 1 | NDC Registry | 53 | Sovereign / UN Open |
| Tier 1 | IISD Earth Negotiations Bulletin | 4 | CC BY-NC-SA 4.0 |
| Tier 2 | Castro et al. 2025 (enb-mining repo) | 1 + scripts | CC BY-NC-SA 4.0 |
| Tier 3 | IPCC AR6 WGII chapters | 4 | IPCC Open Use |
| Tier 3 | COP30 Brazilian Presidency | 6 | Public statement |
| Tier 4 | IMF ND-GAIN CSV | 19/20 CINA | CC BY 4.0 |
| Tier 4 | OWID CO2, PRIMAP-hist, WRI Climate Watch, OECD, CAT, Plano Clima 16 sectoral, Korea MOFA/MOE | 30+ | mixed |
225 manifest entries, 100 % licence + sha256 tracked. Storage ~720 MB raw + 35 MB processed. All collectors live in src/collect/ and re-run idempotently against any free LLM stack.
For each (country, issue) pair, we issue a Stage 1 prompt (docs/04_stage1_stance_extraction.md) that requests strict JSON v1.3:
{
"stance_score": <[-1, 1]>,
"stance_category": "<6-class>",
"key_demands": [...], "red_lines": [...], "flexibility_signals": [...],
"evidence_quotes": [{"quote", "location"}],
"frame_type": "<scientific|justice|sovereignty|security|development|mixed>",
"instrument_signals": {"nodality":[...], "authority":[...], "treasure":[...], "organization":[...]},
"salience_score": <[0, 1]>,
"procedural_signals": {"is_chair_role", "is_pen_holder", "drafts_text_for_issue"},
"confidence": <[0, 1]>
}We run the extraction with k = 5 multi-samples at temperature 0.3, aggregate by confidence-weighted mean, and compute a Beta-binomial 95 % credible interval on the rescaled
Cross-LLM consistency. CINA Stage 1 supports five LLM backends: Gemini 2.5 Flash-Lite, Groq Llama 3.3 70B, Ollama qwen2.5:3b (local), OpenRouter, and Anthropic. We treat each as an independent coder and compute Krippendorff's α at the interval level; this provides a reliability measure that is partially decoupled from shared-model bias because the five providers were trained on different corpora and RLHF pipelines. On the n = 78 country-issue pairs, raw α = 0.876 and bias-corrected α = 0.933 (after per-LLM mean-centering removes systematic offsets, the largest of which is the Ollama qwen2.5:3b under-estimate of −0.158; src/stage1_extract/cross_llm_consistency.py). We position this as a practical lower bound on real-coder reliability, distinct from the simulated-panel α reported in §5.4.
Country, issue, and negotiating-group nodes form a heterogeneous graph; edges are typed across four relations: has_stance, member_of, similar_to, co_chairs. We report two complementary realisations.
(i) NetworkX + igraph + leidenalg. Leiden community detection (Traag, Waltman & van Eck 2019; resolution γ = 1.0) and five centrality measures (PageRank, betweenness, eigenvector, degree, closeness). Cross-issue Apriori-style motif mining (min support 0.2, min issues ≥ 2) detects frame-coherence patterns. Used for §4.1 and §4.2 community-level results.
(ii) Heterogeneous R-GAT (relational graph attention). Implemented in PyTorch with relation-specific linear projections and per-relation attention heads (Schlichtkrull et al. 2018; Veličković et al. 2018), trained jointly on three heads — stance regression, coalition classification, contested-issue prediction — with multi-task loss $\mathcal{L} = \mathcal{L}{\text{stance}} + 0.5\mathcal{L}{\text{coal}} + 0.5\mathcal{L}_{\text{contest}}$. Hyper-parameters: hidden dim 64, 4 attention heads, 2 layers, Adam lr 1e-3 with weight decay 1e-4, 200 epochs, seed 42 (src/stage2_graph/rgat.py). Training converges to validation Spearman ρ = 0.708 ± 0.01 with coalition accuracy 0.77 (10 / 13 countries) and contested P@3 = 0.67 (Figure 8). Crucially, the learned attention weights — without explicit supervision on procedural authority — are dominated by the co_chairs relation (mean attention 1.00) over member_of (0.35), similar_to (0.28), and has_stance (0.08), recovering the procedural-authority signal posited by Tallberg (2010). This emergent attention pattern is reported in §4.2 alongside the Leiden community result.
Briefing generation operates on the Stage 2 JSON plus the deduped evidence-quote pool. Each generated sentence is post-hoc validated against seven rules (numeric claim ↔ analysis JSON, country mention ↔ stance database, structural claim ↔ network metric, etc.); failures are auto-removed and logged. The Korean ministerial template (docs/09) yields nine sections plus four appendices (Evidence Traceability, Data Lineage, Uncertainty, LLM Provider Attribution). A parallel English version targets academic audiences.
CINA was developed under an internal review protocol involving multiple LLM-instantiated reviewer roles (a policy-science reviewer and an IR reviewer) that critiqued each pipeline iteration before quality-gate sign-off; the protocol is documented in docs/10_council_protocol.md for reproducibility but should be understood as an internal development hygiene mechanism, not a substitute for external peer review.
We report three primary empirical observations that follow from the methodology of Section 3, plus two single-case findings (Sections 4.4–4.5) we present as candidate signals warranting further investigation. We deliberately frame these as "observations," not "discoveries," given the limitations enumerated in Section 5.6.
Across the six adaptation issues we coded, GGA-IND scored 6.1 on the Authority sub-axis (Hood 1983; Howlett 2019), the lowest in the corpus, while showing high Nodality usage. This is consistent with the explicit textual posture of FCCC/PA/CMA/2025/L.25E §7 — "voluntary, non-prescriptive, non-punitive, facilitative" — and with §9's negative-authority clause ("shall not create new financial obligations"). We interpret this as a quantitative correlate of the soft-law character of the GGA indicator framework, in line with Howlett's instrument-calibration theory; we do not claim it is the first such quantification, since dictionary-based and supervised approaches to NATO coding have been reviewed by Capano et al. (2025). (Figure 1, top row.)
4.2 Observation 2 — Leiden community detection produces a horizontal cleavage consistent with regime-complex theory; emergent R-GAT attention recovers Tallberg procedural authority
Stage 2 Leiden (Traag, Waltman & van Eck 2019; resolution γ = 1.0) returns two stable communities at modularity 0.31 over n = 13 nodes: C0 = {Brazil, Multi (UAE-Belém), African Group, EU} (development-frame consistent) and C1 = {AOSIS, India, South Korea, LMDC, China} (mixed/justice/sovereignty). C0 unites the G77 chair, the EU HAC anchor, and AGN bridging; C1 unites principled vulnerability advocates with sovereignty defenders. The partition is stable across resolution sweep γ ∈ [0.7, 1.3]. We read this as consistent with the horizontal cleavage component of Keohane & Victor's (2011) regime-complex hypothesis — but we caution against reading too much into a 13-node clustering: the result needs replication on a substantially larger node set before any general claim is warranted. (Figure 5, similarity network.)
A complementary R-GAT analysis (Section 3.3) provides convergent evidence at the edge level. After 200 epochs of multi-task training (validation Spearman ρ = 0.708, coalition accuracy 0.77, contested P@3 = 0.67) on the same graph, the learned attention weights are dominated by the co_chairs relation (mean 1.00) over member_of (0.35), similar_to (0.28), and has_stance (0.08). Because the model receives no procedural-authority supervision, this represents an emergent recovery of Tallberg's (2010) chair-channel hypothesis: chair edges are most informative for stance, coalition, and contested prediction simultaneously. (Figure 8, R-GAT panels (a)-(c).)
We caveat that a Bayesian variance decomposition (3-level: country / formal-group / latent-regime; src/analysis/bayesian_hierarchical.py) on the same data yields a different ranking — country-level random effect explains 54 % of stance variance, formal-group 37 %, latent regime (Leiden community) only 1 %. We disentangle this apparent tension at the modeling level: Leiden modularity measures the connectivity pattern of edge structure (whether countries form internally-cohesive subgraphs in similarity space), whereas the Bayesian σ_regime measures the stance-variance attribution to a community-membership covariate. The two modeling targets are not equivalent — a community partition can be topologically valid (significant Leiden modularity) while contributing little to the variance of node-level stance attributes. This distinction was not made in our v3.0 paper and is added here as a methodological caveat. We additionally report a 1000-run random-graph permutation test (Figure 12, v4.0): the observed Leiden modularity 0.31 has p < 0.001 against random partitions of the same edge density, indicating that the topological signal itself is not noise. Resolution of which interpretation (cleavage as topology vs cleavage as variance) better fits the regime-complex hypothesis requires the longitudinal extension proposed in §7.
4.3 Observation 3 — Stance variance alone correctly identifies the COP30 contested set (small-sample retrospective signal)
Ground truth, from the IISD ENB final report on COP30: the contested-issue set in the adaptation track was {GGA-IND voluntary-vs-mandatory, ADAPT-FIN base-year, L&D-OP contributor expansion}. Predicted top-3 by Stage 1 stance variance alone — that is, with no chair-power feature, no historical interaction frequency from Castro et al. (2025), and no learned weights — was {GGA-IND (var = 0.34), L&D-OP (0.21), ADAPT-FIN (0.15)}. This yields P@3 = R@3 = 1.00 on N = 3 contested issues. The small N matters: this is not a generalisable accuracy claim, but an encouraging single-case signal that stance dispersion captured by the multi-axis schema carries information about future contestation. We will revisit this on COP31 prospectively. (Web-demo §4 evaluation panel.)
Brazil's domestic Plano Clima (16 sectoral plans, 2024–2035) deploys all four NATO axes at an aggregate usage of 67 % (combining Authority, Nodality, Treasure, Organization). The COP30 GGA L.25E text — drafted under Brazilian presidency — collapses to predominantly nodality-based instruments at 48 %, with three "shall not" clauses constituting a negative-authority envelope. We define Δ = IRR_domestic − IRR_international = 0.714 − 0.410 = 0.304 as a single-case quantitative measurement of this divergence. We propose Δ as a candidate metric at the intersection of Two-Level Games (Putnam 1988) and instrument calibration (Howlett 2019; Capano et al. 2025). Since this is a single chair-country in a single year, no claim of theoretical generality is made; replication across COP25–COP30 chair countries is the natural next step. (Figure 2; Brazil-paradox panel of the web demo.)
Comparing the L.25 advance draft and the L.25E final text yields TF-IDF cosine ≥ 0.95 on every aligned paragraph; that is, the chair-circulated draft and the adopted text are essentially identical in this issue. Combined with §7's four-burst hedging (voluntary + non-prescriptive + non-punitive + facilitative in successive sentences), this is consistent with Tallberg's (2010) formula-control channel having been exercised pre-circulation, i.e., the adopted text was already crystallised when the advance text was distributed. This is a single-issue finding and we treat it as a candidate signal of "pre-crystallised formula," not a confirmed pattern; replication on at least three chair-led decisions is needed.
A baseline using only realist features (CO2 per-capita, share of global CO2, log-GDP) on cosine similarity yields F1 = 0.560 against pseudo-truth coalition labels (vs. random F1 = 0.440; Cohen κ = 0.216; McNemar χ² = 16.1, p < 0.0001 vs. random). The realist features alone do significantly better than chance but stay well below the 0.7+ that operational policy use would require. The diagnostic error case — USA–Saudi cosine = 0.889 on realist features, despite the two states sitting in different negotiating blocs — illustrates why frame, instrument, and procedural variables carry orthogonal signal.
| Metric | CINA | Threshold | Status |
|---|---|---|---|
| Spearman ρ | 0.658 | ≥ 0.6 | ✅ |
| MAE | 0.183 | ≤ 0.25 | ✅ |
| 6-class accuracy | 0.59 | — | reasonable |
| 6-class macro F1 | 0.45 | ≥ 0.55 | 🟡 marginal |
n = 34 overlap pairs (CINA × calibration). Bootstrap 95 % CI on ρ = [0.42, 0.83]. Power 0.92. Figure 1 visualises agreement.
ARI proxy = 0.42 vs official negotiation-group labels; modularity 0.31; coalition F1@3 = 0.55. Leiden C0/C1 partition is stable across resolution sweep γ ∈ [0.7, 1.3].
Ground truth: COP30 Belém Adaptation Indicators contested set = {GGA-IND voluntary-vs-mandatory, ADAPT-FIN base-year, L&D-OP contributor expansion} (per IISD ENB final report).
Predicted top-3 by stance variance only: GGA-IND (0.34), L&D-OP (0.21), ADAPT-FIN (0.15) — 3 / 3 correct, P@3 = R@3 = 1.00, F1@3 = 1.00.
This is the strongest single number in the paper: stance dispersion alone, with no chair-power feature, recovers the entire contested set. (Web demo §4 evaluation panel.)
We constructed a panel of five expert personas (a Korean Energy Economics Institute policy researcher, a KAIST IR professor, a Ministry of Foreign Affairs climate negotiator, a Ministry of Environment adaptation officer, and a Global Environmental Politics journal editor) and prompted an LLM (Claude Sonnet 4.5) to score the generated ministerial briefing from each persona's viewpoint on a 5-dimension Likert rubric. The simulated panel yielded a mean of 4.53 / 5 and an inter-persona Krippendorff α of 0.905, with hallucination rate of 1.5 % under post-hoc verification.
We explicitly flag this as a methodological limitation. A simulated-persona panel is not a substitute for real expert evaluation: the same model that generated the brief is, indirectly, evaluating its own output through a persona prompt, which is known to inflate scores and inter-rater agreement (cf. LLM Psychometrics Review 2025, Kearns et al.). The α = 0.905 should therefore be read as an upper bound under shared-model bias, not as a real human-coder reliability measurement. Real-expert validation with at least three external coders (target: KEI, KAIST, MOFA) is queued for a follow-up study.
| Variant | Spearman ρ | MAE | Δ ρ |
|---|---|---|---|
| A0 Full CINA | 0.658 | 0.183 | (base) |
| A1 No graph | 0.625 | 0.187 | −5 % |
| A2 No calibration | 0.592 | 0.201 | −10 % |
| A3 No multi-sample | 0.605 | 0.198 | −8 % |
| A4 No evidence grounding | 0.559 | 0.210 | −15 % |
| A5 No hypergraph | 0.638 | 0.185 | −3 % |
A4 is the largest single contributor. Removing the fuzzy-match evidence-grounding step costs 15 percentage points of Spearman ρ. We read this as empirical support for the design intuition that tying every output claim to a verbatim source quote is what makes a stance-extraction system credible — both to downstream graph analysis and to a human reviewer.
Six limitations bear directly on how the results above should be interpreted:
- Sample size: n = 98 stance records across 13 countries × 6 issues. While this is sufficient for a single-case methodological pilot, generalisable claims require n ≥ 300 with broader country coverage; we treat the present numbers accordingly.
- Single retrospective case (N = 3 contested issues): Task C's P@3 = R@3 = 1.00 is encouraging but reflects one COP cycle and three issues. Prospective COP31 validation is the appropriate next test.
- Single-case Δ = 0.304: The Brazil domestic-international policy-instrument divergence is measured for one chair-state in one year. Replication across COP25–COP30 chair countries is needed before any general claim about chair-mediated translation gaps.
- Simulated expert panel (Task D): As discussed in §5.4, persona-prompted self-evaluation is a known source of inflated agreement. The Krippendorff α = 0.905 is therefore an upper bound under shared-model bias.
- Graph attention not learned: Stage 2 uses NetworkX + igraph + leidenalg. R-GAT is design-level only in the present paper; learned attention weights are future work.
- Calibration set: n = 50 supervised pairs (28 verified + 22 placeholder). External two-coder Krippendorff α has not yet been measured, and Castro et al. (2025) cooperation/conflict ground truth from SWISSUbase is queued but not yet incorporated.
We additionally note potential confirmation drift in the multi-round development protocol (Section 3.5): although the protocol introduces internal critique, the reviewer roles share the LLM provider and may exhibit shared blind spots.
The 30-cell Korea NAP × GGA crosswalk yields IRR_Korea = 0.653 (95 % CI [0.55, 0.71]) — Accept-eligible. The single weakest cell is L&D-OP × 사회·경제 adaptation = 0.39, traceable to Korea's EIG-membership-plus-middle-income-contributor dual identity ambiguity (MOFA press release seq=376685, 2025-12-08).
We therefore recommend, for the COP31 Türkiye negotiation:
- Activate Korea's NAP pen-holder status (Stage 1 extraction confirms
pen_holder = truefor Korea on NAPs) by inputting the 3-tier (national-province-municipal) NAP model into the Belém-Addis 2-year work programme. - Internationalise the JT-Korean model: Carbon Neutrality Framework Act §50 (vulnerable-population protection) is a deliverable for the COP31 plenary statement and a Track-B paper.
- Issue an L&D-OP institutional support pledge of USD 5–10 M, claiming a Korean board seat at FRLD on operational-efficiency grounds.
- Articulate EIG dual identity by aligning with Switzerland and bridging to AILAC via Mexico.
- Lead the GCF operational-efficiency agenda to deflect contributor-base-expansion pressure.
Track A submission package: ministerial_briefing_ko.md, paper.md (this), IRR_Korea.md, korean_nap_gga_crosswalk.csv, six 300-dpi figures, evaluation_report.md. Submitted to the Kookmin Graduate Global Climate Leadership course (May 2026).
We presented CINA, a multi-axis LLM stance-extraction pipeline coupled with Leiden community detection and graph-grounded briefing generation, and piloted it on the COP30 adaptation negotiation cycle. Within the limitations of a single-case methodological pilot (Section 5.6), three observations carry the paper: (a) Leiden community detection on the multi-axis stance tensor produced a horizontal cleavage that is consistent with Keohane & Victor's (2011) regime-complex hypothesis; (b) stance variance alone correctly identified the 3-of-3 COP30 contested issues in retrospective testing; and (c) we propose Δ, a single-issue domestic-international policy-instrument divergence metric, with a first measurement of Δ = 0.304 for Brazil at COP30. The framework's main methodological contribution is the joint extraction of stance + instrument + frame + procedural signals, and the graph-grounded generation discipline whose ablation accounts for the largest single component effect (−15 % Spearman ρ).
We see four priority extensions, in order of urgency: (i) prospective application to COP31 (Türkiye, November 2026) under the falsifiable hypotheses pre-registered in PREREGISTRATION_COP31.md with code freeze 2026-09-01; (ii) external two-coder Krippendorff measurement using KEI / KAIST / MOFA reviewers in place of the simulated panel reported in Section 5.4 — to be deployed alongside the cross-LLM α = 0.93 reliability bound established here; (iii) longitudinal extension across COP21-COP30 with the difference-in-differences / synthetic-control / IV identification strategy specified in CAUSAL_IDENTIFICATION_STRATEGY.md, testing whether the single-case Brazil Δ = 0.304 generalises causally; and (iv) deepening of the R-GAT layer with Schlichtkrull-style relation-specific embeddings and explicit interpretability probing of the chair-edge attention dominance reported in §4.2.
Source code: github.com/zxsa0716/cina (MIT). Documentation, briefings, and crosswalks: CC BY 4.0. All third-party data licences are tracked in data/manifest/manifest.jsonl. We invite replication and welcome external validation, particularly from external expert coders working in the climate-diplomacy domain.
- Bayer, P., & Urpelainen, J. (2013). External sources of clean technology: Evidence from the Clean Development Mechanism. Review of International Studies, 39(4).
- Capano, G., Howlett, M., & co-authors (2025). Applying Hood's NATO framework to quantitative text analysis in policy studies: Theory, methods, and empirical applications. Working paper / preprint. ResearchGate 400309939.
- Castro, P., Kristof, V., Kammerer, M., & Cogne, T. (2025a). Participation, cooperation and conflict in UN climate negotiations. Scientific Data. DOI: 10.1038/s41597-025-06262-4.
- Castro, P., et al. (2025b). Dynamic networks of negotiation for international climate change cooperation. Environmental Sociology. https://doi.org/10.1080/23251042.2025.2507287.
- Finnemore, M., & Sikkink, K. (1998). International norm dynamics and political change. International Organization, 52(4), 887–917.
- GIZ Data Lab (2025). NegotiateCOP: an AI prototype for equal climate negotiations. https://negotiatecop.org/about (accessed May 2026).
- Goh, E. (2007). Great powers and Southeast Asian regional security strategies: omni-enmeshment, balancing, and hierarchical order. RSIS Working Paper.
- Haas, P. M. (1992). Introduction: epistemic communities and international policy coordination. International Organization, 46(1), 1–35.
- Hochstetler, K., & Milkoreit, M. (2014). Emerging powers in the climate negotiations: shifting identity conceptions. Politics & Policy.
- Hood, C. (1983). The Tools of Government. London: Macmillan.
- Howlett, M. (2019). Designing Public Policies: Principles and Instruments (2nd ed.). Routledge.
- IISD Earth Negotiations Bulletin (2025). Belém Climate Change Conference Summary, Vol. 12, No. 888.
- Keohane, R. O., & Victor, D. G. (2011). The regime complex for climate change. Perspectives on Politics, 9(1), 7–23.
- Platt, J. (1999). Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers.
- Putnam, R. D. (1988). Diplomacy and domestic politics: the logic of two-level games. International Organization, 42(3), 427–460.
- Sebenius, J. K. (1983). Negotiation arithmetic: adding and subtracting issues and parties. International Organization, 37(2), 281–316.
- Snow, D. A., & Benford, R. D. (1988). Ideology, frame resonance, and participant mobilization. International Social Movement Research, 1, 197–217.
- Steinberg, R. H. (2002). In the shadow of law or power? Consensus-based bargaining and outcomes in the GATT/WTO. International Organization, 56(2), 339–374.
- Tallberg, J. (2010). The power of the chair: formal leadership in international cooperation. International Studies Quarterly, 54(1), 241–265.
- Tollison, R. D., & Willett, T. D. (1979). An economic theory of mutually advantageous issue linkages. International Organization, 33(4), 425–449.
- Traag, V. A., Waltman, L., & van Eck, N. J. (2019). From Louvain to Leiden: guaranteeing well-connected communities. Scientific Reports, 9, 5233.
- UNFCCC (2025a). FCCC/PA/CMA/2025/L.25E — Belém Adaptation Indicators decision.
- UNFCCC (2025b). FCCC/PA/CMA/2025/L.24 — Tripling adaptation finance decision.
- Vaccari, S., et al. (2025). Tracing inclusivity at UNFCCC conferences through side events and interest group dynamics. Nature Climate Change. DOI: 10.1038/s41558-025-02254-9.
- Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., & Bengio, Y. (2018). Graph attention networks. International Conference on Learning Representations (ICLR).
- Schlichtkrull, M., Kipf, T. N., Bloem, P., van den Berg, R., Titov, I., & Welling, M. (2018). Modeling relational data with graph convolutional networks. Extended Semantic Web Conference (ESWC).
- Abadie, A., Diamond, A., & Hainmueller, J. (2010). Synthetic control methods for comparative case studies. Journal of the American Statistical Association, 105(490), 493–505.
- VanderWeele, T. J., & Ding, P. (2017). Sensitivity analysis in observational research: introducing the E-value. Annals of Internal Medicine, 167(4), 268–274.
- Zhang, T., et al. (2025). AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N. Proceedings of Machine Learning Research (PMLR), v267.
본 연구는 UNFCCC 협상 텍스트의 다축 stance 추출(stance + NATO 4축 정책수단 + 5 frame + 절차 권한 신호)을 단일 LLM 추출 단계에서 수행하고, 그 산출물을 Leiden 커뮤니티 검출과 graph-grounded 브리핑 생성으로 연결하는 방법론적 파이프라인 CINA를 제안한다. 본고는 단일 사례에 대한 회고적 검증을 보고하는 방법론 예비(pilot) 논문이며, 일반화 가능한 결론을 주장하지 않는다. n = 98 stance records(13개국 × 6 이슈)에 기반한 4-task 평가 결과, Spearman ρ = 0.658(전문가 reference 대비, Task A), COP30 contested 이슈 3/3 정확 예측(N = 3, Task C, 작은 표본의 고무적 신호), simulated 5인 페르소나 패널 평균 4.53 / 5(Task D — 본 패널은 LLM 시뮬레이션이며 실제 외부 전문가 검증은 향후 과제로 명시한다). 단일 사례 관찰로서 브라질의 국내(Plano Clima) ↔ 국제(L.25E) 정책수단 사용률 차이를 Δ = 0.304로 정량화하여, Two-Level Games(Putnam 1988)와 정책수단 calibration(Howlett 2019; Capano 등 2025)의 교차점에 위치한 후보 메트릭으로 제안한다. 한국 적응정책의 30-cell IRR은 0.653(95 % CI [0.55, 0.71])이며 L&D-OP 영역이 가장 낮은 0.39로, COP31 협상을 위한 5건의 정책 권고를 도출한다. 코드 MIT, 문서 CC BY 4.0, github.com/zxsa0716/cina.
Manuscript prepared 2026-05 by Heedo Choi, Kookmin University Department of Climate Technology Convergence. Single-author graduate research project; not yet peer-reviewed.