RECON: A Recipe-Driven, Evidence-Fused, Neuro-Symbolic Multi-Agent Architecture for Autonomous Document Compliance Auditing across Heterogeneous Enterprise Repositories

<title>Abstract</title> <p>Document compliance review in regulated organisations stays slow, uneven and hard to defend, because human reviewers disagree on borderline clauses and rarely record a verifiable account of why a document passed or failed (Morande &amp; Shah, 2020; Lipton, 2018). Rule-only engines read precisely yet miss paraphrase, while pipelines built only on a Large Language Model (LLM) read flexibly yet hallucinate and drift between runs (Ji et al., 2023; Gao et al., 2023). This paper presents RECON, a Recipe-Driven, Evidence-Fused, neuro-symbolic multi-agent architecture for autonomous document compliance auditing across heterogeneous enterprise repositories. A compliance regime is declared as a YAML recipe rather than written as program code, so a control owner adds a regime by authoring a file. During an audit a five-agent crew is configured from the selected recipe, and findings from three independent signals (deterministic symbolic rules, fuzzy near-miss detection and agentic LLM reasoning over hybrid retrieval) are merged by union and reduced by a rule identifier. Each verdict carries a confidence interval (CI) whose half-width narrows as corpus size and finding count rise, and each run is written to an append-only ledger chained by a SHA-256 hash so that later edits break verification. A reference build was evaluated on a mixed corpus of 43 enterprise documents that spanned contracts, treasury records, policy manuals and regulatory returns. The fused pipeline produced 57 findings and a failing verdict of 30 of 100 with a reported interval of [28, 32], where a focused contract batch returned a warning verdict of 72 of 100 over 13 findings. An ablation that removed the symbolic signal lowered precision, and a removal of the agentic signal lowered recall. The results suggest that pairing symbolic precision with neural recall, under a tamper-evident record and an evidence-aware interval, may support compliance decisions that are reproducible and defensible, warranting cautious generalisation beyond the present corpus.</p>