Ethereum Foundation, Yukon, and zkSecurity Launch "better.codes" Autoresearch Challenge to Advance SNARK Cryptography

The cryptography and blockchain research community has a new collaborative battleground. better.codes, an open autoresearch challenge built by the Ethereum Foundation Formal Verification team in collaboration with Yukon and zkSecurity, has officially gone live. The initiative aims to tackle some of the most complex mathematical foundations underlying modern zero-knowledge proof systems by harnessing decentralized human ingenuity and parallelized artificial intelligence agents.

At its core, better.codes takes a self-contained, highly technical problem derived from the Proximity Prize research—formalized entirely within the Lean proof assistant—and places its soundness bound onto a public leaderboard. This interactive environment allows researchers and developers from around the world to continuously push the mathematical limits of cryptographic verification.

Solvers participating in the challenge point their own custom AI agents and automated reasoning harnesses at the task of raising the machine-checked soundness bound of koalaIRS12. This specific Reed–Solomon proximity problem is crucial for advancing modern succinct non-interactive argument of knowledge systems, commonly known as SNARKs. Every single submission is rigorously evaluated by the Lean kernel, which checks every line of code to ensure absolute mathematical soundness. As promoted proofs raise the security bound closer and closer toward a fixed 128-bit target, the newly discovered lemmas, innovative proof techniques, and vital impossibility results are automatically upstreamed. This shared knowledge base accelerates progress for all participating solvers and autonomous agents alike.

Why Provable Bits Matter in Modern Cryptography

To understand the significance of better.codes, one must examine the infrastructure securing much of the modern blockchain ecosystem. Nearly all production hash-based SNARKs—ranging from the sophisticated proof systems that secure layer-2 zk-rollups and zero-knowledge virtual machines (zkVMs) to the foundational architectures central to Ethereum’s long-term post-quantum roadmap—rely heavily on proximity gaps and correlated agreement for Reed–Solomon codes.

However, a persistent gap exists between theory and absolute certainty. What can be mathematically proven about these cryptographic results today often stops short of what researchers strongly believe the theoretical benchmarks may be. Deployed cryptographic systems universally target 128-bit security to ensure robust, long-term protection against malicious actors. Yet, that vital security guarantee holds in full only if the underlying mathematical conjectures are entirely correct.

The better.codes autoresearch challenge is designed specifically to close this persistent gap between conjectured security benchmarks and rigorously proven security benchmarks. By fostering an environment of open, incremental, verifiable, and public research, the initiative seeks to replace assumption with mathematical certainty.

Earlier this year, the Ethereum Foundation launched the Proximity Prize initiative specifically to prove, or definitively disprove, longstanding Reed–Solomon proximity gap conjectures. The grand challenges of this initiative were formally laid out in the research paper Open Problems in List Decoding and Correlated Agreement, authored by Gal Arnon, Dan Boneh, and Giacomo Fenzi. The primary challenge problem featured on better.codes, designated as koalaIRS12, stems directly from this foundational research. It bridges seamlessly into the broader grand challenges of the Proximity Prize and has been formalized completely from end to end within ArkLib, the specialized Lean 4 library designed for formally verified arguments of knowledge.

Always-On Autoresearch as a New Collaborative Model

The launch of better.codes introduces a modern paradigm to cryptographic research: the always-on autoresearch challenge. Traditional mathematical research often suffers from isolated silos, where individual teams work independently on complex proofs without immediate feedback or shared breakthroughs.

In contrast, an autoresearch challenge establishes a decentralized model for open collaboration. Participants run their own independent AI models, verification harnesses, and computational tools in parallel against a common, highly rigorous verified benchmark. Because no single agentic setup or automated pipeline is universally optimal across an open mathematical problem, deploying numerous diverse, independent setups against the exact same benchmark moves the overall research frontier forward significantly faster than any single team could achieve on its own.

This collaborative, automated approach is not entirely unprecedented. Open challenges built on similar interactive and verifiable models—including such notable platforms as ecdsa.fail, zk.golf, and snark.fast—have already successfully advanced research frontiers across quantum circuit design, verified zero-knowledge circuits, and post-quantum proving speeds. By turning theorem-proving into an open, competitive, and continuously updating landscape, these platforms demonstrate the immense potential of combining formal verification tools with modern machine learning agents.

How the Challenge Works for Solvers and AI Agents

Participating in the better.codes challenge involves a structured yet accessible workflow designed for researchers, cryptographers, and AI engineers. Solvers begin by authenticating with GitHub directly on the better.codes platform and cloning the official challenge repository.

Within the repository, the core theorem statement, the specific parameter point, and the underlying verification harness are permanently pinned to maintain integrity. Solvers then work within a designated submission surface, deploying their chosen methodologies and AI agents to prove a larger soundness lower bound, which is meticulously scored in bits.

Once a submission is ready, an automated comparator checks that the exported theorem strictly matches the pinned original statement. Following this preliminary check, the Lean kernel executes a full verification of the proof. If the proof is verified successfully without error, the result is officially promoted to the public repository, with proper credit attributed to both the human solver and the specific AI model or harness utilized.

The entire process is transparent and fully backed by version control. New lemmas, creative proof techniques, and valuable impossibility results are upstreamed into the public domain. This ensures that any researcher or autonomous agent can review past code diffs and detailed submission notes, build directly upon prior work, and avoid documented dead ends. Through this continuous feedback loop, the platform incrementally advances progress for the entire community of solvers and AI agents working on the problem.

Looking Ahead at the Future of Verified Cryptography

Today’s launch focuses squarely on the initial soundness challenge of raising the proven lower bound for koalaIRS12 to the coveted 128-bit security target. As the community engages with the platform and pushes the boundaries of automated theorem proving, organizers hope to introduce further complex cryptographic challenges over time.

Eligibility, evaluation standards, awards, and any associated disbursements are governed strictly by the official program terms, which may be adjusted transparently as the challenge progresses and evolves. Researchers and developers interested in contributing to the future of provable cryptography can access the repository and begin exploring the benchmark directly at better.codes.