Neuro-Symbolic Financial Reasoning via Deterministic Fact Ledgers
VeNRA is a verifiable reasoning agent for financial AI that eliminates hallucinations by grounding every computation in typed, auditable fact ledgers rather than learned associations. The architecture demonstrates that LLM outputs can be made fully traceable and deterministic without sacrificing analytical flexibility. This work directly informs Fact AI Lab's approach to eval-first AI systems in regulated financial environments, where every inference must be explainable and verifiable.
VeNRA (Verifiable Numerical Reasoning Agent) addresses a foundational problem in deploying LLMs for financial analysis: uncontrolled hallucination in numerical and logical reasoning.
The Problem
Standard LLM-based financial assistants generate plausible-sounding numbers that may be mathematically inconsistent, sourced from the wrong time period, or simply invented. In regulated environments - asset management, financial reporting, compliance review - this is not a tolerable failure mode.
Approach
VeNRA introduces a Universal Fact Ledger: a typed, append-only data structure that records all values the system may reason over, along with their provenance (source document, page, timestamp, extraction confidence). When the LLM reasons, it does not invent values - it retrieves them from the ledger by typed key.
The reasoning layer is a deterministic Python executor. Arithmetic, aggregation, and conditional logic are executed as Python code generated by the LLM, then run under a restricted interpreter. The LLM provides the logic; the interpreter provides correctness guarantees.
Verification: Every answer includes a full audit trail: which ledger entries were accessed, which Python operations were applied, and what the intermediate values were. A separate verifier checks the final answer against the audit trail.
Results
On a held-out set of financial Q&A pairs constructed from real earnings documents and regulatory filings, VeNRA achieves near-zero hallucination on numerical claims while maintaining answer quality comparable to standard RAG pipelines on factual recall.
The system is currently being integrated into Fact AI Lab's compliance review tooling.