GreenRisk v0.1.0 — Explainable Greenwashing-Risk Scoring (Public Release)

First public release of GreenRisk: an explainable greenwashing-risk scorer
for corporate climate disclosures. Given one paragraph of text, it returns
a 0–100 risk score, the exact rule-activation trace that produced it, and a
W3C PROV-O provenance graph binding the result to the pinned model
revisions and instrument version used.

What's in this release

• The scoring instrument, frozen at tag rulebase-locked-v1: five
  pinned ClimateBERT classifiers feeding a 17-rule Mamdani fuzzy inference
  system. See docs/architecture.md for the full
  pipeline and design rationale.
• Two independent layers of validation evidence, both run against the
  frozen instrument (see docs/validation.md):
  • Large-scale: 1,009 TCFD disclosure paragraphs, convergent validity
    against a published cheap-talk baseline (Spearman ρ = 0.60), and
    discriminant validity within the most vague quartile (ρ = 0.43 between
    commitment and risk, where the baseline alone sees no variation).
  • Held-out, case-level: 15 real, regulator-adjudicated cases (DWS,
    Volkswagen, HSBC vs. Microsoft, Ørsted), scored once against the
    instrument after it was frozen. In-scope greenwashing separates from
    rigorous disclosure at AUC = 0.87; the predicted out-of-scope misses
    were confirmed 4/4.
• A documented, precisely bounded limitation: the instrument detects
  vagueness-based ("cheap talk") greenwashing; it is not designed to catch
  greenwashing built on selective or corrupted specific numbers. Both
  failure modes are demonstrated on real cases, not hypothesized.
• Full reproducibility: pinned model revisions, committed artifacts
  (artifacts/), and exact commands in MASTER_PLAN.md
  to regenerate every figure and validation result from scratch.

Licensing and citation

Released under Apache License 2.0. See
docs/acknowledgements.md for citations to the
ClimateBERT models and papers this project builds on, and
CITATION.cff if you use this work — GitHub renders a
"Cite this repository" button from it.

Known limits of this release

• Validation layer 2 is intentionally small (n = 15) and held-out by
  design, not a large benchmark — reported as a demonstration with a large
  effect size, not a statistically definitive claim on its own.
• Scope is paragraph-level, TCFD-trained signals; not tested outside that
  disclosure framework.