What Is the Harmonic Aristotle API? Formal Verification AI for Developers

Most AI models guess. They predict the next token, synthesize plausible answers, and sometimes get things wrong in ways that are hard to detect. For general conversation, that is an acceptable tradeoff. For mathematics, protocol logic, or smart contract code, it is not.

The Harmonic Aristotle API is built around a different approach. Instead of returning a probabilistic answer, Aristotle proves its answers are correct using Lean 4, an open-source formal proof language. If the proof fails, the system returns no result. That architecture makes it one of the few AI systems where developers can treat the output as mathematically verified rather than merely plausible.

This article explains what the Aristotle API is, how it works technically, who is using it today, and why its formal verification approach is increasingly relevant to developers building on blockchain infrastructure where correctness is not optional.

Aristotle is an AI system built by Harmonic, a Palo Alto AI lab co-founded by Vlad Tenev (also co-founder of Robinhood) and Tudor Achim. The company describes its mission as building Mathematical Superintelligence — an approach to AI that prioritizes rigorous, verifiable reasoning over statistical pattern matching.

The Aristotle API gives developers programmatic access to this system via HTTPS. Once authenticated with an API key, developers can submit Lean 4 proof files with incomplete proof steps (marked with sorry placeholders), plain English math problems, or entire research papers in LaTeX format. Aristotle then attempts to complete, prove, and formally verify the submitted content.

The Python SDK — available as aristotlelib on PyPI — wraps the API into a command-line tool that works with local Lean project directories. A few representative commands illustrate the scope:

The agent can run autonomously for up to 24 hours without human intervention on larger proof tasks. According to Harmonic, Aristotle ranked first in formal mathematics according to the ProofBench evaluation by ValsAI, ahead of the closest competitor by 15 percentage points.

The key distinction between Aristotle and a general-purpose AI is that it produces output in Lean 4, a programming language designed for writing mathematical proofs that can be mechanically checked by a computer. When a Lean proof compiles, it means every step in the reasoning chain has been verified by a deterministic kernel — not by another AI, and not by a human.

Harmonic's CEO Tudor Achim explained to TechCrunch that within the quantitative reasoning domains Aristotle supports, it genuinely guarantees no hallucinations. This is possible because the verification step is entirely separate from the generation step. Aristotle may draft a proof using informal reasoning, but the output is only accepted if Lean's type checker validates every step deterministically.

A VentureBeat analysis of Lean4 describes the approach clearly: each step in the AI's chain of thought translates a claim into Lean 4's formal language and a proof is required. If the proof fails, the system knows the reasoning was flawed. This step-by-step audit trail eliminates the class of errors where an AI produces a confident but wrong answer.

Aristotle's technical architecture combines three main components:

Proof search. A Monte Carlo Graph Search algorithm explores the space of possible Lean 4 tactics, guided by a large transformer model used as both a policy and value function. The system targets proof bottlenecks first by prioritizing the most challenging subgoals.

Lemma-based informal reasoning. Aristotle generates informal proofs in natural language, breaks them into intermediate lemmas, formalizes each lemma into Lean 4, and iterates based on the formal feedback loop. This allows the system to leverage the much larger corpus of mathematical knowledge available in natural language while still requiring machine-verified output.

Test-time training. For difficult problems, Aristotle can retrain itself on search traces from its own attempts during inference, improving its performance on a specific problem without changing the base model.

The system proved five of six problems at the 2025 International Mathematical Olympiad — the most prestigious mathematics competition in the world — earning a gold-medal-equivalent score. What distinguishes Aristotle's performance from OpenAI's and Google's IMO results is that its solutions were formally verified in Lean 4, not just correct natural-language answers.

Harmonic has raised significant capital on a rapid timeline:

Paradigm's participation in the Series B is notable for developers in the blockchain space. Paradigm is one of the most active institutional investors in crypto infrastructure, and its backing of Harmonic reflects the growing connection between formal verification research and onchain systems.

Since its public launch, mathematicians and quantitative researchers have adopted Aristotle to accelerate proofs and explore new research directions. The system has been used to solve multiple long-standing open Erdős problems that had been unsolved for decades, and its code contributions have been accepted into Mathlib — the main community library of formalized mathematics for Lean — with no modifications required.

Formal verification has been standard practice in safety-critical domains — aviation firmware, medical devices, nuclear control systems — for decades. It is becoming relevant to blockchain development for similar reasons: the cost of errors is very high, and human review does not scale reliably with system complexity.

Smart contract exploits cost the DeFi ecosystem an estimated $3.8 billion between 2024 and 2025. A reentrancy vulnerability drained $47 million from a major DeFi protocol in March 2025 in under 90 seconds — despite the contract having been audited by three separate firms. The core problem is not auditor competence. It is that human review, even a thorough human review, does not produce mathematical guarantees.

Formal verification does. A property that has been expressed as a Lean 4 theorem and compiled successfully is a mathematical guarantee, not a risk estimate.

Mistral AI's release of Leanstral in March 2026 — an open-source Lean 4 proof agent — signals that the market for formal verification tooling is expanding. Multiple teams now treat smart contract verification via proof assistants as a practical engineering requirement, not a research exercise.

The Aristotle API accelerates this by making Lean 4 proof generation accessible through a simple HTTP interface. Developers do not need to be experts in formal methods to submit verification tasks. They write their protocol logic, describe the properties they want to guarantee, and Aristotle attempts to construct a machine-checked proof that those properties hold.

The most direct application for teams building on-chain is using Aristotle to verify that smart contracts satisfy stated properties. A developer can encode contract invariants as Lean theorems — for example, that a liquidity pool's total asset value is conserved across rebalancing operations — and submit them to the Aristotle API. If Aristotle constructs a valid proof, those properties are guaranteed by mathematical verification rather than testing.

Research from PropertyGPT, an AI-powered formal verification tool that received a Distinguished Paper Award at NDSS 2025, demonstrates that LLMs can transfer knowledge from existing human-written properties to generate security specifications for new contracts at scale. Aristotle provides the underlying proof engine that validates whether those specifications hold.

Many research-grade blockchain protocol papers include informal mathematical arguments that claim certain security properties hold. Translating those arguments into Lean 4 and formally verifying them have historically required significant expertise. Aristotle changed that workflow.

A developer or researcher can write a LaTeX description of a protocol's security model, submit it to the Aristotle API using the aristotle formalize command, and receive a Lean 4 file that encodes the argument and attempts to prove it. Where proofs succeed, the security claim is formally established. Where they fail, the system identifies exactly where the reasoning breaks down — which is precisely the kind of gap that traditional auditing misses.

This matters for protocol developers thinking about programmable money movement. When money movement is encoded in smart contract logic that reacts to market conditions, verifying that the conditional execution paths are exhaustive and correct is a mathematical problem, not just a software testing problem.

Aristotle can also generate counterexamples: given a statement that turns out to be false, it produces a concrete execution trace demonstrating why. For protocol developers, this is useful for debugging invariants during design. A developer who believes a liquidity invariant holds can submit it to the API — if Aristotle produces a counterexample, they have a specific case that breaks their assumption before deployment.

The Verina benchmark, developed by researchers at UC Berkeley and Meta, evaluates verifiable code generation — producing code alongside a formal proof of its correctness. Harmonic reports that Aristotle achieved state-of-the-art performance on Verina, demonstrating that the formal reasoning capabilities that make it effective for mathematical proof also transfer to software correctness.

For teams building infrastructure that handles financial operations — routing, settlement, treasury management — the ability to generate code accompanied by machine-checked correctness proofs is a meaningful step toward high-assurance software development at production scale.

Access is available through the Aristotle API portal, where developers can sign up, log in, and generate API keys from the dashboard. The Python SDK requires Python 3.10 or later and installs through pip or uv. Authentication is handled by setting an environment variable:

bash

From there, the CLI allows direct submission of Lean projects, LaTeX papers, and natural-language prompts. For teams that want to integrate the API into CI/CD workflows — running formal verification checks on every protocol commit — the --wait flag blocks until the proof attempt completes, making it compatible with standard pipeline tooling.

Harmonic also runs a $1 million Research Grant Program, offering Principal Investigator Awards for researchers leading ambitious formalization projects and Rising Mathematician Awards for developers and students exploring the limits of formal AI reasoning.

The Aristotle API sits at the intersection of two trends that are both gaining traction in 2026: AI systems that operate as autonomous agents in production environments, and the growing demand for provably correct behavior in financial and protocol contexts.

For teams building stablecoin infrastructure — where cross-chain routing, atomic settlement, and programmable execution need to behave predictably under a wide range of conditions — formal verification provides a stronger guarantee than any combination of testing and auditing. It answers a different question: not "did we test for this case?" but "can we prove this holds for all cases?"

Aristotle is not a replacement for good engineering practice. It is a tool that lets developers elevate certain properties — correctness guarantees that matter most — to mathematical certainty. As Harmonic's research note on cryptographic proof systems indicates, Aristotle has already made contributions to CompPoly, a library of machine-checked mathematics for the polynomial computations behind modern zero-knowledge proof systems. That is a sign that the system's capabilities are penetrating the cryptographic foundations of blockchain infrastructure.

Understanding how networked stablecoin movement depends on protocol-level correctness at every hop clarifies the stakes. A routing system that moves value across a dozen chains operates over combinatorially many paths. Formally verified invariants at the contract layer are what make that kind of system trustworthy at scale.

The Aristotle API makes those invariants achievable — not just in academic settings, but as a practical developer workflow.

What is the Harmonic Aristotle API? The Aristotle API is a developer interface to Harmonic's formal reasoning AI. It accepts Lean 4 proof files, LaTeX math papers, and natural-language problems, then attempts to formally verify or prove them using Lean 4's machine-checkable proof language. Outputs are only returned when a valid proof has been constructed and verified.

What is formal verification and why does it matter? Formal verification is the process of mathematically proving that a system or piece of code satisfies a stated property. Unlike testing — which checks specific cases — a formal proof covers all possible cases. In blockchain and DeFi contexts, formal verification provides stronger guarantees than auditing because it is not subject to human error or incomplete coverage.

How is Aristotle different from standard AI coding assistants? Standard AI coding assistants generate code probabilistically and can produce incorrect outputs that look correct. Aristotle generates Lean 4 proofs and verifies them before returning results. If the proof fails to compile, no answer is returned. Within the domains it supports, this eliminates hallucinations as a category of error.

What programming language does Aristotle use for verification? Aristotle uses Lean 4, an open-source functional programming language and proof assistant developed by Microsoft Research and now maintained by the Lean Focused Research Organization. Lean 4 has a large community library called Mathlib that contains thousands of formalized mathematical theorems.

Who is building Harmonic? Harmonic was co-founded by Tudor Achim (CEO) and Vlad Tenev, who is also co-founder and CEO of Robinhood. The company has raised $295 million across three funding rounds from Sequoia Capital, Kleiner Perkins, Paradigm, Ribbit Capital, Index Ventures, and others.

What are the main use cases for the Aristotle API? Key use cases include: filling sorry placeholders in Lean 4 proof files, formalizing LaTeX math papers into Lean, generating counterexamples for false mathematical statements, verifying smart contract properties, and producing formally verified code for high-assurance software development.

How do I get an API key for Aristotle? Sign up or log in at aristotle.harmonic.fun, navigate to Dashboard → API Keys, and create a new key. The Python SDK (aristotlelib) is available via PyPI and requires Python 3.10 or later.