RECON: A Recipe-Driven, Evidence-Fused, Neuro-Symbolic Multi-Agent Architecture for Autonomous Document Compliance Auditing across Heterogeneous Enterprise Repositories
This paper introduces RECON, a neuro-symbolic multi-agent architecture that combines deterministic rules, fuzzy detection, and LLM reasoning with a YAML-driven recipe system and tamper-evident auditing to deliver reproducible, defensible, and evidence-fused document compliance reviews across heterogeneous enterprise repositories.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to solve a massive jigsaw puzzle, but the pieces are scattered across different rooms, some are written in invisible ink, and some look like they fit but are actually slightly the wrong shape. This is the daily reality for companies trying to follow complex rules, like laws about data privacy or financial safety. They have to check thousands of documents—contracts, bank statements, and policy manuals—to make sure they aren't breaking any rules.
To do this, they usually rely on two types of helpers. The first is a "Rule-Follower," a strict computer program that looks for exact words. It's like a robot that only knows the word "apple" and will miss a picture of a red fruit if you call it a "crisp, round fruit." The second helper is a "Big Brain," a type of artificial intelligence that understands language and meaning. It's great at seeing that a "crisp, round fruit" is an apple, but sometimes it gets confused, makes things up, or changes its mind every time you ask it the same question. The big challenge in science right now is figuring out how to combine these two helpers so they can work together without fighting, creating a system that is both smart enough to understand tricky language and strict enough to be trusted by regulators.
This is where a new system called RECON comes in. Think of RECON as a super-team of five specialized agents working together to audit documents, designed to be both flexible and unbreakable. Instead of writing complex computer code to set up new rules, a human in charge just writes a simple "recipe" file (like a cooking recipe, but for rules). This recipe tells the team exactly what to look for.
When the team starts an audit, they don't just guess. They use three different ways to find problems, kind of like using a metal detector, a magnifying glass, and a detective's intuition all at once. First, the "Rule-Follower" scans for exact matches. Second, a "Fuzzy Detective" looks for phrases that are almost the right words, catching things that are slightly different but still suspicious. Third, the "Big Brain" reads the context to understand the meaning. The magic happens when they combine their notes. They merge all their findings into one list, but they are careful not to count the same problem twice. If the strict robot says "No problem" but the fuzzy detective and the big brain both say "This looks risky," the team trusts the evidence that is strongest.
The paper reports that when they tested this system on a mixed pile of 43 real-world business documents (including contracts and financial reports), it found 57 different issues. It was incredibly accurate, getting the right answer 91% of the time (precision) while also catching 88% of the actual problems (recall). This is a big deal because previous systems usually had to choose: be very strict and miss hidden problems, or be very flexible and make mistakes. RECON suggests you can have both.
To make sure no one can cheat or change the results later, the system writes every single step of its work into a digital "ledger" that is chained together with a special lock (a SHA-256 hash). If someone tries to sneakily erase a mistake or change a verdict, the whole chain breaks, and the tampering is immediately obvious. It's like a diary where every page is glued to the next; if you rip one out, the whole book falls apart.
The researchers also added a "confidence meter" to every verdict. If the system checks only a few documents, it says, "I'm pretty sure, but the answer could wiggle a bit." If it checks hundreds of documents, the answer becomes very tight and precise. In their tests, when looking at a small batch of contracts, the system gave a warning with a wide range of uncertainty, but when looking at the full set of 43 documents, it gave a specific failing score with a very narrow range of certainty.
The study suggests that this mix of strict rules, fuzzy matching, and smart AI, all recorded in a tamper-proof log, could make checking documents much faster and more trustworthy. However, the authors are careful to say this is based on a pilot test with a specific set of banking documents. While the results look promising, they suggest that more testing is needed to see if it works just as well for every type of company and every kind of rule. They also point out that the system still needs a human to write the initial "recipes," so it's not fully automatic yet, but it does give humans a powerful new tool to do their jobs better.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.