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 are inherently n-ary (binding reporting entities, base years, target horizons, greenhouse gas scopes, and verification standards), while comparative multi-entity queries suffer from representation collapse, where verbose filings monopolize retrieval budgets. 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 high-dimensional corporate data as an n-ary hypergraph, capturing multi-constraint relations without combinatorial edge explosion. Second, we introduce the Claims as Hyperedges formalism in a dedicated discourse graph layer, where claims connect directly to regulatory audit requirement nodes; deterministic credibility and greenwashing risk scores are computed directly via requirement node coverage. Third, we implement Company-Stratified Fair Retrieval, ensuring exact parity and 100% entity recall in comparative assessments across hundreds of reporting organizations.
In empirical evaluations across a verified automotive enterprise corpus spanning 589 companies and 110 multi-category 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 demonstrate that pure learned neural re-rankers degrade recall by 15.2 points due to the erasure of exact-match and schema anchors, establishing that contrastive representations must serve as a calibrated residual signal (β = 0.15) to structural heuristics.
Key Innovations
01 n-ary Hypergraphs
Captures multi-constraint corporate commitments across entities, scopes, base/target years, and assurance standards without pairwise edge explosion.
02 Claims as Hyperedges
Formalizes claims as hyperedges over regulatory requirement nodes. Yields deterministic credibility and greenwashing risk scores from node coverage.
03 Stratified Fair Retrieval
Prevents representation collapse in comparative queries by allocating per-entity top-k quotas, guaranteeing 100% entity recall across hundreds of companies.
04 Calibrated Hybrid Scoring
Proves contrastive representations must serve as a residual signal (β = 0.15) alongside exact lexical anchors; avoids the 15.2-point collapse of pure neural re-rankers.
05 Coordinate Provenance
Every retrieved passage, claim, and table cell ties deterministically to document MD5 hash, page number, and bounding-box coordinates.
Empirical Benchmark Scale
BibTeX Citation
@article{moka2026claims,
title={Claims as Hyperedges: Auditable Multi-Hop Retrieval and Discourse Reasoning over Enterprise Disclosures},
author={Moka, Varun},
journal={arXiv preprint},
year={2026},
month={October},
url={https://varunmoka.com/paper}
}