GraphLLM

Abstract

Climate-risk review requires connecting evidence about a bank’s exposures with evidence about its response. Those connections are difficult to recover when disclosures differ in entity boundary, portfolio coverage, reporting period and level of detail. We develop GraphLLM, a framework combining schema-constrained language-model extraction with a typed knowledge graph and a Source–Addressed–Remaining (SAR) diagnostic. Source and response claims are represented separately, with their provenance and epistemic status preserved. A per-dimension correspondence relation assesses whether response evidence addresses an exposure cluster through identification, governance, measurement, targets, assurance and disclosure quality. The result is a decomposable set of unresolved evidence requirements rather than a composite firm score. Five discovery channels integrate sectoral, geographical, counterparty, public-claim and scenario evidence within the same representation. We specify the graph operations, extraction and verification architecture, and worked mechanisms for transition, physical and liability risk. A proposed Lloyds Banking Group FY2025 pilot separates extraction accuracy, diagnostic validity, external corroboration and robustness, with explicit comparison and false-negative audits. The paper’s contribution is methodological: it makes exposure–response reasoning inspectable and empirically testable. Completed benchmark results and evidence of cross-firm generalisation are not reported in this version.

Correspondence: Tian Han · University of Bristol