Prove
Formal
A proof shows that a precise model of the code keeps a stated property for every input the model allows, not only the inputs a test tried, and a computer checks every step.
Experimental See the example →
Plain words
Basics
A test tries the inputs someone picked. A proof covers every input the model allows.
- Property
- The rule you need, such as "the total paid never passes the cap".
- Model
- A precise description of the code, or of its design, that the proof is about.
- Proof
- A step-by-step argument that a small checker program accepts.
Honest scope
Limits
A proof covers the property as written and the model as built, under the assumptions it states.
- A property that misses the point gives a sound proof of the wrong thing.
- Anything outside the model, such as the compiler, the chip, other services or the people calling the code, is assumed, not proven.
- A proof about the design says nothing about whether the code follows it.
- In Anthropic's words, “Formally verified” does not mean “bug-free.” Source
Recent advances
Frontier
-
Crash-free libraries
Claude wrote proofs, checked by the Rocq proof checker, that 239 entry points in 123 open-source libraries never crash or fail a check on any input under 64 MB, counting the memory errors a standard detector catches. The proofs assume one thread on a model of a RISC-V computer running Linux; hundreds of bugs were found and fixed on the way.
Anthropic and Theorem → -
Fermat proof
Claude, working largely on its own for 11 days, wrote a complete proof of Fermat's Last Theorem, building on Lean's shared maths library, that the Lean proof checker accepts. It checks a known proof; it is not new mathematics.
Anthropic → -
AI Composer
Certora released an early, open-source tool that checks AI-written smart contract code against safety rules with its Prover while the code is being written.
Certora → -
AI specs
Informal Systems had an AI update a Quint model of a consensus protocol, checked the model by hand and by machine, then had the AI write the code and tests from it, with people guiding the design. One case study; the authors say the process has much to mature.
Quint → -
AlphaProof
Google DeepMind published how AlphaProof learns to find Lean proofs by practising on millions of problems. With AlphaGeometry 2 it reached a silver-medal score at the 2024 International Mathematical Olympiad.
Nature → -
Aristotle
Harmonic's Aristotle gave Lean-checked solutions to five of the six 2025 International Mathematical Olympiad problems, a gold-medal-equivalent result.
Harmonic → -
Checked answers
Automated Reasoning checks in Amazon Bedrock became generally available. They use mathematical logic to check that an AI model's answers follow a set of written rules.
AWS → -
Halmos invariants
Halmos 0.3.0 from a16z checks that a rule, written as a Solidity test, holds across every sequence of smart contract calls up to a set length (two calls by default).
a16z → -
Kontrol CI
Runtime Verification used Kontrol to prove that Optimism's bridge contracts on Ethereum stop processing withdrawals while paused, and added the proofs to Optimism's automatic checks.
Runtime Verification →
Worked example
ProvedFunding Cap
The Wallet Bench pays each agent it tests a small amount of USDC, and ETH for fees. This model describes how those payments stay within each run's total while several tests run at once, a failed try is retried, or a payment's result on Base is not known yet.
- Property
- For each token, the payments the bench keeps, and the money that really moves on Base, never pass the run's total.
- Model
- The design of the funding step, written in Quint, not the code itself. It assumes Base never undoes a finalized block, and that a Base server that is behind shows an older record, never a false one.
- Proof
- Apalache proves both rules hold for runs of any length within the model's limits: three tests funded at once, a total of two payments of each token, and one retry each.
- Found
- Two ways past the total. If the database ran at repeatable read, three tests could each keep a payment against a total of two; a test on Postgres shows the same, and Techtree runs at read committed. And a Base server that is behind could make a lost payment look replaced, freeing room past the total; Techtree now reads Base at its finalized block, and the model was checked again.
- Code
- The proof covers the design, not the code. A test on Postgres and tests against a stand-in Base server tie the two together.
Submit
Submit Code
- The code, or a link to its repository.
- The property you need guaranteed, in plain words.
- Where it runs, such as a chain, a server or a device.
We reply by email with a proposed scope.