Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

Aaron Sterling

Home/Archive/arXiv

arXiv · 2606.04903

Provably Auditable and Safe LLM Agents from Human-Authored Ontologies

Aaron Sterling

View original source ↗Read source document ↗Export metadata ↓

Abstract

We introduce the LLM agent architecture Agentic Redux, intended for use with nontrivial problem domains that require linear auditability. Using the typed lambda calculus, we prove that, run on appropriate domains, Agentic Redux executions are semantically guaranteed to be correct, with all decisions recorded in an append-only ledger. We present two production-grade appropriate domains, in healthcare billing compliance, and security vulnerability disclosure. Working code for Agentic Redux run on both domains is available in a supporting code repository. We also introduce Ontology-First Agent Design, a methodology for creation of agent frameworks on a problem domain, in which a human expert ontologizes the problem domain with Basic Formal Ontology, and then assigns an LLM to derive roles that agents and humans-in-the-loop can fill, in order to work the problems in the domain.

Explore related subjects

cs.LO cs.AI cs.MA cs.PL

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Aaron Sterling. 2026-06-03. Provably Auditable and Safe LLM Agents from Human-Authored Ontologies. https://arxiv.org/abs/2606.04903

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

PICID: Proof-Driven Clause Learning in Neural Network Verification

Current Deep Neural Network (DNN) verifiers are typically designed to prioritize scalability over reliability. Reliability can be reinforced through the generation of proofs that are checkable by trusted, external proof checkers. To date, only a handful of verifiers support proof production; and these rely on verifier-specific formats, and balance between scalability, proof detail, and the trustworthiness of their proof checker. In this tool paper, we introduce PICID, a DNN verifier that produces proofs in the standard Alethe format for SMT solving, checkable by an independent checker. PICID implements a parallel CDCL(T) architecture that integrates the state-of-the-art, proof-producing CaDiCaL SAT solver with the Marabou DNN verifier. Furthermore, PICID leverages UNSAT proofs to derive conflict clauses. Our evaluation shows that PICID generates valid proofs in the vast majority of cases and significantly outperforms existing tools that produce comparable proofs.

cs.LO↗

arXiv2026

Confluence of conditional rewriting modulo

Sets of equations E play an important computational role in rewriting-based systems R. The equivalence relation =E induced by E introduces a partition of terms into E-equivalence classes on which rewriting computations, denoted ->R/E and called rewriting modulo E, are issued. This paper investigates confluence of ->R/E, usually called E-confluence, for conditional rewriting-based systems, where rewriting steps are determined by conditional rules. We rely on Jouannaud and Kirchner's framework to investigate confluence of an abstract relation R modulo an abstract equivalence relation E on a set A. We show how to particularize such a framework to be used with conditional systems. Then, we show how to define appropriate finite sets of conditional pairs to prove and disprove E-confluence. We introduce (i) Logic-based Conditional Critical Pairs, which do not require the use of (often infinitely many) E-unifiers to provide a finite representation of the local peaks considered in the abstract framework. We also introduce (ii) parametric Conditional Variable Pairs which are essential to deal with conditional rules in the analysis of E-confluence. Finally, we introduce (iii) Down Conditional Pairs which are often necessary to disprove E-confluence. Our results apply to well-known classes of rewriting-based systems, improving on previous results. As for unconditional systems, our results apply to Equational Term Rewriting Systems, first investigated by Huet and then by Jouannaud, and Jouannaud and Kirchner, among others. As for conditional systems, our results also apply to conditional rewrite theories and Maude.

cs.LO↗

arXiv2026

Verifying Numerical Methods with Isabelle/HOL

Modern machine learning pipelines are built on numerical algorithms. Reliable numerical methods are thus a prerequisite for trustworthy machine learning and cyber-physical systems. Therefore, we contribute a framework for verified numerical methods in Isabelle/HOL based on ITrees. Our user-friendly specification language enables the direct declaration of numerical programs that can be annotated with variants and invariants for reasoning about correctness specifications. The generated verification conditions can be discharged via automated proof methods and lemmas from the HOL-Analysis library. The ITrees foundation interacts with Isabelle's code generator to export source code. This provides an end-to-end path from formal specifications with machine-checked guarantees to executable sources. We illustrate the process of modelling numerical methods and demonstrate the effectiveness of the verification by focusing on two well-known methods, the bisection method and the fixed-point iteration method. We also contribute crucial extensions to the libraries of formalised mathematics required for this objective: higher-order derivatives and Taylor's theorem in Peano form. Finally, we qualitatively evaluate the use of the framework for verifying numerical methods.

cs.LO↗