When an AI writes you code, it produces something that looks right, passes your tests, gets reviewed, and ships. But passing tests isn't the same as being correct — especially when AI is handling both the code generation and the tests. That gap is where bugs live.
And the gap is getting harder to see. AI is getting better at writing code that runs flawlessly on the first try. But "it works" doesn't mean "it won't break."
lemmafit adds an external check to the AI coding loop. Instead of just generating code and declaring it done, the agent has to produce logic that passes a mathematical verification step. Not more tests, not layers of AI checking AI. An external proof. If the logic is wrong, it doesn't pass.
With lemmafit, the agent's role is no longer to generate and judge a final output. Instead, its role is to propose implementation candidates that are externally checked by a deterministic process.
Lemmafit plugs into your AI coding agent (currently only Claude Code) and guides it to back up its logic — counters, totals, balances, conditions — with guarantees. At the end, you get statements like: "the number of items in a shopping cart can never be negative." That guarantee holds for every edge case, forever. Not because it was tested or reviewed 100 times. Because it was automatically proven.
Works as an npm package. Zero config. You don't have to know anything about formal verification or proofs to use it. You just prompt Claude and lemmafit handles the rest.
Version 0.1.0:
Works with Claude Code
Catches logic errors that tests miss
Claude keeps you updated on verification progress
Open source
