Why it mattersA team held back by the cost of a wrong answer has a new kind of tool to try. The gate is a proof, not a test.
Companies hold the data to answer a business question. They often will not let an AI agent answer it. An answer off by a decimal place is worse than no answer. Axiomat launched on Y Combinator this week with a plan to move that gate. The YC page says Axiomat converts a business rule and an English question into a mathematical definition. It then runs the Lean proof assistant to check that the query it would run matches that definition. The YC page states the rule: "If we can't prove it, the query doesn't run." The launch had 6 votes in its first hours.
What the product does
Axiomat gives a user a plain English interface in front of their data. The YC page says the system builds a definition of the question from the user's words. It builds a second definition of the business rule from the company's rules. Lean, the proof assistant, decides whether a given SQL query answers exactly that question under those rules. A query that passes runs. One that fails is rejected. The system translates the proven query back to plain English, so the user sees the question the system actually answered.
Who runs it
The founders are Ethan Kharitonov, Jack Naimer and Elad Perlman. The YC post is from Y Combinator's Fall 2026 batch. The post does not list employers or prior projects for the three founders. It gives no revenue figure, no round size, and no named customers. The company's own site is at axiomat.com.
What the YC post does not say
The claim is a strong one for anyone who has built an AI agent on top of a database. A query cannot run unless it has a proof. Four things the YC post does not state, which a buyer would ask. The time a proof takes on a real business question. The size of the Lean specification a buyer has to write for their own rules. The queries the system refuses to handle, because not every business question has a short Lean definition. And how the system behaves when the Lean specification is itself wrong, which has happened on published Lean work before.
What this changes for a team building software: a product that would otherwise be blocked by a "wrong answer is unacceptable" rule now has a vendor to test. The test is a shortlist of real business questions on real data, measured on two things: how many of the questions the system refuses, and how many of the ones it answers are right. A pilot with those two numbers is the only real test. It is the way to know whether the proof gate holds for the buyer.
Source
- Axiomat launch page, Y Combinator
- Axiomat website
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
