solvi: the model proposes, code decides — verifiable business decisions in Python

⚡ The gist in 5 seconds - The idea: a language model only proposes facts, while the final decision is made by deterministic code. - Where it's available: solvi is an open-source Python library under the Apache-2.0 license. - Limitation: free-form text answers are not supported — only typed questions and verifiable facts. ### 🔍 What was found The author Exactor introduced solvi — a library built on the principle of "the model proposes, code decides." The architecture splits the work: models extract facts from documents (quotes, selections from a list), while ordinary Python code checks them against types, hard rules, and quote offsets. Every step is recorded in a trace chained with hashes, so a decision can be reproduced six months later and presented to an auditor. In version 0.5.0, Python type annotations became the contract. They describe facts, questions, and allowed answers — including new variants: "not stated in the text," an exact text excerpt, ranking, and a number with an interval. If the model proposes a quote that doesn't literally match the document, it is rejected (grounding). A choice outside the declared options is also rejected (closed set), and when confidence is low, the question abstains. Hard validation can set an answer regardless of any confidence level. The article gives an example: a taxi expense report is solved twice with a single catalog — once by code alone and once with models involved. ### 💡 Why it matters For business processes that require accountability to auditors and strict rule compliance, directly using LLMs is risky. A model told "lawsuit threat — high priority" responded to a message containing the word "lawyer" with "queue: billing" at 0.94 probability and "priority: normal" at 0.49. It computes numbers by similarity, not arithmetically. solvi solves this by ensuring the model never gets the final say: its proposals pass code checks, and rules are guaranteed to be enforced. ### 🧩 Context The author arrived at solvi after experimenting with hypercomplex numbers (quaternions, octonions) — they only provided an advantage where the task is covered by an exact law. He then studied the "deciding" models Jev and Laya: they are fast, but there's no way to verify why they answered the way they did, and they can override a rule or invent a quote. The conclusion: the most useful part is the computable, not the learnable. That's why in solvi the model remains an advisor, while decisions are made by verifiable code.

🔗 Read on habr.com