GreenRisk

Explainable greenwashing-risk scoring for corporate climate disclosures: pinned ClimateBERT signals → a Mamdani fuzzy rule base → an auditable 0–100 risk score with a W3C PROV-O provenance graph.

Every score ships with an activation trace showing exactly which rules fired and why, and a provenance graph binding the result to the exact model revisions and instrument version that produced it. Nothing is a black box.

Instrument status: frozen at tag rulebase-locked-v1.

Documentation

• docs/architecture.md — the full pipeline: the climate-relevance gate, the four signals, the fuzzy rule base, and the design rationale behind it.
• docs/validation.md — the validity evidence: a large-scale statistical comparison against a published baseline, and a held-out test against real, regulator-adjudicated cases, including the instrument's documented scope limitation.
• docs/acknowledgements.md — citations for the ClimateBERT models, datasets, and standards this project builds on.
• MASTER_PLAN.md — the repository map, the locked instrument at a glance, and exact reproduction commands.

Repository Map

• models.py, linguistic_variables.py, rule_base.py — locked scoring instrument.
• main.py — small CLI for scoring already-computed signal probabilities.
• scripts/ — artifact producers for figures, full-corpus runs, contrast-set runs, baselines, and provenance.
• scripts/validation/ — validation harnesses for anchors, hash behavior, and integrated scoring.
• data/contrast_set.csv — held-out contrast set used after the lock.
• artifacts/ — committed outputs used by the validity argument.
• docs/ — public documentation (architecture, validation, acknowledgements).

Quick Start

uv sync
uv run python main.py --specificity 0.20 --commitment 0.90 --sentiment-asymmetry 0.40 --netzero 0.70
uv run python scripts/validation/integration_seam_test.py -n 20 --gate 0.5

The CLI path does not load ClimateBERT; it scores four precomputed signal values through the locked Mamdani rule base and prints the rule-activation trace. Scripts that call models.py may download/load the pinned Hugging Face model revisions on first run. Inference uses a GPU automatically if one is available and falls back to CPU otherwise — a GPU is not required.

Citation

If you use GreenRisk, please cite it using CITATION.cff (GitHub renders a "Cite this repository" button from this file), or via its Zenodo archive: 10.5281/zenodo.21122389 — this concept DOI always resolves to the latest release; see the Zenodo record for version-specific DOIs (e.g. v0.1.0 is 10.5281/zenodo.21122390). See docs/acknowledgements.md for the papers and models this project is built on.

License

Apache License 2.0 — see LICENSE. The underlying ClimateBERT models are also Apache 2.0 licensed; see docs/acknowledgements.md for full citations.