Claims as Hyperedges: Auditable Multi-Hop Retrieval and Discourse Reasoning over Enterprise Disclosures

Abstract

Retrieval-Augmented Generation (RAG) and pairwise Knowledge Graph (KG) systems have emerged as prominent frameworks for synthesizing unstructured enterprise documents. However, regulated compliance and sustainability disclosures present structural hurdles that current paradigms fail to address: corporate commitments bind multiple facts at once (reporting entities, base years, target horizons, greenhouse gas scopes, and verification standards), while comparative multi-company queries suffer from representation collapse, where lengthy filings monopolize search results. Furthermore, prevailing claim verification methods rely on subjective large language model (LLM) sentiment rather than deterministic audit criteria.

In this work, we propose HyperGraphRAG, a 3-layer architecture for verifiable information retrieval and discourse analysis over complex corporate disclosures. First, we model multi-dimensional corporate data as a hypergraph, capturing multi-part commitments in a single connection without fragmenting facts. Second, we introduce the Claims as Hyperedges approach in a dedicated discourse graph layer, where claims connect directly to regulatory audit checklist items; objective credibility and greenwashing risk scores are computed from how many required proof points a claim satisfies. Third, we implement Company-Stratified Fair Retrieval, ensuring every company gets equal representation and 100% coverage in side-by-side comparisons across hundreds of reporting organizations. In empirical evaluations across a verified automotive enterprise corpus spanning 589 companies and 110 domain queries, HyperGraphRAG improves overall Recall@10 from 87.05% to 87.98% and lifts comparative query recall by +3.3 percentage points (to 62.74%). Crucially, we show that relying purely on neural AI re-rankers drops recall by 15.2 points because they blur exact numbers and dates—proving that AI similarity works best as a light 15% boost alongside exact keyword and structural checks.

Five Architectural Pillars

• 1
  Multi-Fact Hypergraph Memory (Layer 2): A climate claim is never just two things—it links a company, an emissions scope, a baseline year, a target year, a number, and an audit standard all at once. Traditional graphs split these into separate pairs and lose the context; our hypergraph keeps all six pieces locked together as a single unit.
• 2
  Checklist-Based Claim Verification (Layer 3): Instead of asking an AI chatbot whether a company's claim "sounds believable," we link every claim directly to a checklist of required regulatory proof points (baseline, methodology, third-party assurance, scope coverage). Credibility and greenwashing risk scores are calculated directly from which proof points are present or missing.
• 3
  Fair Side-by-Side Company Comparisons: When comparing multiple companies, standard search engines let companies with 300-page reports crowd out companies with 20-page reports. We reserve an equal share of search slots for every company in the query, guaranteeing 100% company coverage on every comparison.
• 4
  Exact Numbers First, AI Similarity Second: Pure AI search often confuses "Scope 1 in 2025" with "Scope 3 in 2030" because the sentences sound similar. We anchor search on exact keywords, table structures, and regulatory terms (85% weight), using neural AI similarity as a 15% tie-breaker.
• 5
  Page & Highlight-Box Proof (Layer 1): Every extracted number, table cell, and claim is permanently tied to the original PDF's digital fingerprint, exact page number, and highlight coordinates—so an auditor can click any claim and inspect the original source on the page.

Citation (BibTeX)

If you build on this research or reference HyperGraphRAG, please cite:

@article{moka2026hypergraphrag,
  title     = {Claims as Hyperedges: Auditable Multi-Hop Retrieval and Discourse Reasoning over Enterprise Disclosures},
  author    = {Moka, Varun},
  journal   = {arXiv preprint},
  archivePrefix = {arXiv},
  eprint    = {2026.gct.hypergraphrag},
  primaryClass = {cs.AI},
  year      = {2026},
  month     = {October},
  note      = {GoCarbonTracker Research, https://gocarbontracker.net/paper}
}