VeNRA: The Architecture Behind Zero-Hallucination Financial AI

A financial AI that is 99% accurate is operationally useless. The 1% doesn't average out — it compounds. A single hallucinated number in a SEC filing analysis, an audit report, or a risk model isn't a minor inconvenience. It's a compliance failure, a liability, and potentially a regulatory event.

This is the problem VeNRA was designed to solve. Not to make AI "more accurate" in the benchmark sense, but to make it verifiable — where every numerical claim in the output traces back to a specific source and every calculation can be independently reproduced.

Why Standard RAG Doesn't Work for Financial AI

The dominant paradigm for grounding LLM outputs is retrieval-augmented generation: retrieve relevant document chunks, pass them to the model, let the model synthesize an answer. This works well for qualitative questions. It breaks for quantitative financial analysis.

The root issue is architectural. When you ask an LLM to "analyze the revenue trend across these 10-K filings," it does two things simultaneously: it retrieves and paraphrases numerical content, and it performs implicit arithmetic. Language models are not reliable calculators. They'll confidently write "revenue grew 23% year-over-year" when the actual figure is 21.7% — because 23% is a more fluent number in that sentence context.

You can't prompt your way out of this. The model's fundamental job is to generate plausible text. Precision arithmetic and source attribution are orthogonal to that objective.

VeNRA's Three Core Components

Universal Fact Ledger (UFL)

Instead of retrieving document chunks, VeNRA retrieves typed, structured facts. A chunk might say "Revenue for Q3 2024 was $4.2 billion, up from $3.9 billion in Q3 2023." The UFL representation is:

revenue_q3_2024: {value: 4.2, unit: "USD_billions", source: "10-K_2024, p.47, line 12"}
revenue_q3_2023: {value: 3.9, unit: "USD_billions", source: "10-K_2023, p.44, line 8"}

The LLM never touches raw numbers. It reasons about typed facts with explicit source attributions attached.

Double-Lock Grounding

Before any numerical claim enters the output, it passes two verification gates:

1. Semantic grounding: does this fact exist in the Universal Fact Ledger? If the model tries to cite a number that isn't in the ledger, it's flagged.
2. Arithmetic verification: if this number is derived (e.g., "revenue grew 7.7%"), can it be reproduced by deterministic Python execution from the ledger values? If not, the claim is rejected.

The derived-calculation check is the most important piece. Most financial hallucinations aren't invented facts — they're plausible but incorrect calculations. Double-Lock catches these before they reach the output.

VeNRA Sentinel

A compact SLM (3 billion parameters) that runs forensic audits on reasoning traces with sub-100ms latency. The Sentinel checks whether the final output is internally consistent with the ledger entries — not just whether individual claims are grounded, but whether the overall narrative coheres with the underlying facts.

In our evaluations, a 3B Sentinel outperforms Gemini-2.5-Flash on financial error detection tasks. Size isn't the bottleneck; the right training objective is.

The Core Insight

The architectural lesson generalizes beyond finance: separate what LLMs are good at from what they're bad at, and give each category to the right system.

LLMs are good at: understanding intent, structuring analysis, synthesizing narrative, handling ambiguity, generating well-formed text.

LLMs are bad at: precise arithmetic, source attribution, consistency over long contexts, detecting their own errors.

VeNRA gives the language parts to the LLM and the precision parts to deterministic systems. The result isn't a smarter LLM — it's a system where the LLM's strengths are preserved and its failure modes are structurally blocked.

Full architecture, evaluation methodology, benchmark results, and implementation details: VeNRA project page.