Skip to content

CHANGELOG.md

Latest commit

 

History

History
70 lines (52 loc) · 5.84 KB

CHANGELOG.md

File metadata and controls

70 lines (52 loc) · 5.84 KB
Raw

Changelog

All notable changes to the CINA project are documented here. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[3.0.0] — 2026-05-05

Added

  • Master end-to-end orchestrator src/run_all.py — single command runs the entire pipeline (S0 environment → S1 stance corpus → S2 R-GAT → S3 briefing → S4 Phase 5 → S5 cross-LLM α → S6 Bayesian decomp → S7 figures → S8 RUN_REPORT.md), with --live-llm, --skip-rgat, --skip-figures flags.
  • LLM provider production smoke test src/stage1_extract/llm_smoke_test.py — validates Gemini, Groq, Anthropic, Ollama, OpenRouter on a CINA-shaped prompt; --offline mode for CI; production-tier judgement (1 OK invokable, 2 OK ensemble, 3+ robust α).
  • Publication-grade figure suite src/viz/publication_figures.py — 10 figures under unified CINA visual identity (300 dpi, navy/primary/accent/warm/success palette, diverging stance colormap). New: fig9 cross-LLM α, fig10 Bayesian decomposition.
  • R-GAT advancement (E1) src/stage2_graph/rgat.py — heterogeneous PyTorch R-GAT with relation-specific attention heads, multi-task training (stance + coalition + contested). 200 epochs reach validation Spearman ρ = 0.71; emergent chair-edge attention 1.00 (vs similarity 0.28) without procedural-authority supervision.
  • Cross-LLM consistency (E2) src/stage1_extract/cross_llm_consistency.py — interval Krippendorff α across 5 providers; per-LLM bias diagnosis; disagreement-driven active-learning sampler. Raw α = 0.876 / bias-corrected α = 0.933.
  • Bayesian variance decomposition (E3) src/analysis/bayesian_hierarchical.py — PyMC 3-level (country / formal-group / latent-regime) + NumPy ML fallback. σ_country = 0.27 (54%), σ_group = 0.22 (37%), σ_regime = 0.04 (1.4%).
  • OSF-style pre-registration (E4) PREREGISTRATION_COP31.md — 4 falsifiable hypotheses with Bonferroni-corrected α = 0.0125 and code freeze planned 2026-09-01.
  • Causal identification strategy (E6) CAUSAL_IDENTIFICATION_STRATEGY.md — DiD with parallel-trends test, synthetic control with placebo permutation, IV using regional rotation timing.
  • Methodology advancement roadmap METHODOLOGY_ADVANCEMENT_ROADMAP.md — 8 specific extension paths with priority matrix and 6-12 month timeline.
  • Self-critical literature review CRITICAL_REVIEW.md — comparison against Castro et al. 2025 (Nature SciData + Environmental Sociology), NegotiateCOP (GIZ 2025), RICE-N (PMLR 2025), Capano et al. 2025 NATO text-analysis review, Vaccari et al. 2025 inclusivity paper.
  • Production-grade README with centred badges, anchor navigation, 3 contribution table, LLM setup matrix, 8-theory integration table.
  • GitHub repository scaffolding: CITATION.cff, .github/FUNDING.yml, issue templates (bug_report.yml, academic_question.yml, config.yml), pull request template, GitHub Actions CI (.github/workflows/ci.yml) and tag-triggered release workflow (release.yml).
  • Community files: CONTRIBUTING.md, CODE_OF_CONDUCT.md, SECURITY.md, CHANGELOG.md (this file).
  • GitHub setup guide GITHUB_REPO_GUIDE.md for one-time UI configuration steps.

Changed

  • All _v2, _v3 versioned filenames consolidated to canonical names (e.g., paper_draft_v3_combined.md → paper.md, IRR_Brazil_2025_v2_negAuth.md → IRR_Brazil.md, evaluation_report_v2.md → evaluation_report.md).
  • Paper v3 abstract, introduction, §3 methodology, §4 observations, §5 limitations, §7 conclusion rewritten under honesty-first framing: "novel" replacing "first", "simulated panel" disclosed in Task D, R-GAT moved from "implemented" to "design + actual implementation now", "8 publishable findings" reframed as "3 primary observations + 2 single-case findings".
  • README badges updated to reflect actual measured values (Spearman 0.658, P@3 1.00, Cross-LLM α 0.93, R-GAT chair attention 1.00).
  • ALL_OUTPUTS_INDEX.md now reflects 10 publication figures, 4 web pages, 5 advancement documents, master pipeline reference.

Fixed

  • fig4_centrality and fig5_similarity_network rewritten without networkx (pure numpy power-iteration PageRank, in-house Fruchterman-Reingold layout) — robust on environments where networkx spectral routines hang.
  • README citation block: removed DOI claim (graduate research project, not peer-reviewed).
  • Internal "Combined Rubric 4.76/5" badge removed (internal metric, not externally meaningful).

Documentation

  • V3_RELEASE_NOTES.md — full release announcement.
  • RUN_REPORT.md — auto-generated by master pipeline with per-step metrics, durations, artefacts, headline takeaways.
  • All paper figure references (fig1–fig10) cross-linked in paper.md, web index, and outputs page.

[2.0.0] — 2026-04-30

Added

  • 5-LLM provider abstraction (Gemini, Groq, Ollama, OpenRouter, Anthropic).
  • Stage 1–3 pipeline with multi-axis stance extraction (NATO 4-axis, frame typology, procedural authority).
  • Phase 5 4-task evaluation (Spearman ρ, ARI, P@K, simulated 5-persona panel).
  • 8 quantitative validation reports (IRR_Korea, IRR_Brazil, L25 formula evidence, realist baseline, AILAC NES, etc.).
  • Korean and English ministerial briefings.
  • 7 paper figures (heatmap, procedural authority, frame consistency, centrality, similarity network, hedging, Brazil paradox).
  • 4-page interactive web demo (index + methodology + visualizations + outputs).
  • 225-entry data manifest with sha256 + license tracking.
  • Multi-agent council protocol for development-time critique.

[1.0.0] — 2026-04-15

Added

  • Initial framework specification (15 methodology documents).
  • Stance schema v1.3 (stance_score + Bayesian CI + NATO + frame + procedural).
  • Data collection scaffolding (UNFCCC, NDC, ENB, IPCC).
  • Country / issue / group identifiers and reference tables.