Raising machine-checked security benchmarks to advance hash-based SNARKs through agentic collaboration
The launch of better.codes serves as a critical infrastructure project for the broader cryptographic community. As the ecosystem pivots toward increasingly complex systems like zero-knowledge rollups (zkRollups), zkVMs, and post-quantum cryptographic primitives, the reliance on high-performance proof systems has intensified. At the heart of these systems lies the Reed-Solomon proximity problem—a fundamental mathematical challenge that determines the soundness of proof systems.
The Problem of Provable Security
In the current landscape of digital infrastructure, most production hash-based Succinct Non-Interactive Arguments of Knowledge (SNARKs) rely on specific conjectures regarding Reed-Solomon codes. These systems are designed to provide a security target of 128 bits, a standard widely considered sufficient to protect against current and near-future computational threats. However, there is a recurring discrepancy in the field: the security level is often "conjectured" rather than "proven."
The security guarantee for these systems holds only if the underlying mathematical conjectures are true. If an attacker discovers a loophole in these assumptions, the entire architecture of a zkRollup or a zkVM could be compromised. The better.codes initiative is designed to close this gap by formalizing these proofs in Lean, an interactive theorem prover, and crowdsourcing the effort to raise the proven soundness bound to the target of 128 bits.
The challenge, titled "koalaIRS12," is the first major milestone for the platform. By pinning this problem to a public, machine-verifiable leaderboard, the project organizers are shifting the research methodology from isolated academic silos to a dynamic, collaborative environment.
Chronology and the Proximity Prize Context
The origins of better.codes can be traced back to early 2024, when the Ethereum Foundation launched the Proximity Prize initiative. This program was established to address the "Open Problems in List Decoding and Correlated Agreement," a seminal research paper authored by cryptographic experts Gal Arnon, Dan Boneh, and Giacomo Fenzi.
The research paper laid out the grand challenges necessary to advance the state of the art in list decoding—a process essential for the error-correction codes used in SNARKs. By formalizing these problems in ArkLib, a dedicated library for formally verified arguments of knowledge in Lean 4, the Ethereum Foundation provided the raw materials for researchers to begin a systematic attack on these proofs.
The development timeline accelerated throughout the year as the Formal Verification team at the Ethereum Foundation joined forces with Yukon and zkSecurity. The goal was to build a platform that could sustain an "always-on" research cycle. Unlike traditional grant-based research, which often results in static, infrequent publications, the better.codes platform allows for incremental progress. Each time a solver submits a proof that raises the soundness bound, the new lemmas and techniques are integrated into the repository. This creates a "ratchet effect," where the floor of proven knowledge is constantly rising, and subsequent researchers can build upon the verified work of their predecessors.
The Mechanics of Autoresearch
The "autoresearch" model pioneered by better.codes is a departure from conventional peer review. It acknowledges that no single human researcher or monolithic AI agent possesses the optimal approach to solve highly complex, multi-dimensional mathematical problems. Instead, the platform invites participants to deploy a diverse array of AI agents, scripts, and manual proofs to tackle the koalaIRS12 problem.
Participants sign in via GitHub and clone the challenge repository, which contains the pinned theorem statement and the verification harness. The workflow is rigorous:
- Submission: Solvers submit their proof, which must be written in Lean.
- Verification: The Lean kernel acts as an immutable judge. It performs a formal verification of the code, ensuring that the logic is sound and the soundness bound has indeed been increased.
- Promotion: Once verified, the proof is added to the repository. The solver is credited, as is the specific AI model or methodology used to generate the result.
- Upstreaming: All successful submissions, including the underlying lemmas and impossibility results, are made public. This allows other participants to identify "dead ends" and pivot their own agents toward more promising strategies.
This transparency is a hallmark of the platform. By making the entire history of attempts available—including the diffs and notes—the project effectively creates a shared intelligence pool. This methodology has been successfully applied in previous niche challenges, such as ecdsa.fail, zk.golf, and snark.fast, which were instrumental in pushing the boundaries of quantum circuit design and post-quantum proving speeds.
Broader Implications for Cryptographic Integrity
The implications of successfully reaching the 128-bit target for the koalaIRS12 problem are significant. First and foremost, it bolsters the foundational integrity of the Ethereum roadmap. As the network transitions to more scalable, privacy-preserving layers, the reliance on provably secure SNARKs will only grow. If the underlying security parameters are mathematically verified rather than just conjectured, the risk of catastrophic failure due to unforeseen mathematical attacks is drastically reduced.
Furthermore, this model suggests a future for academic research that is far more efficient than the current system. Traditional mathematics and cryptography research often suffer from slow feedback loops. By digitizing and formalizing the research process, the better.codes platform allows for a "continuous integration" style of scientific discovery.
The use of Lean 4 is particularly noteworthy. As a proof assistant, Lean provides a high degree of confidence that is simply not possible with traditional prose-based proofs. Because the proofs are machine-checked, they are immune to the subtle logical errors that often plague complex mathematical papers. The collaboration between the Ethereum Foundation and specialized security firms like zkSecurity signals a growing industry-wide recognition that formal verification is not just a theoretical luxury, but a necessity for the survival of decentralized financial infrastructure.
The Path Forward
While the immediate objective is to achieve the 128-bit soundness bound for the koalaIRS12 problem, the organizers view this as merely the first step. The architecture of the platform is designed to be extensible. Once the current challenge is satisfied, it is expected that new, more complex problems will be introduced.
The eligibility and reward structures for the challenge are currently governed by the program’s terms and conditions, which are designed to evolve alongside the project. Participants are encouraged to monitor the better.codes website for updates on scoring, awards, and the introduction of subsequent challenges.
The success of this project will likely serve as a blueprint for other high-stakes technological domains. By turning scientific research into a competitive, collaborative, and verifiable game, the organizers of better.codes are not just solving a specific problem in Reed-Solomon codes; they are stress-testing a new paradigm for human-machine cooperation. In this new era of autoresearch, the barriers to progress are no longer the limits of a single laboratory or a single team, but the collective ability of a global community to iterate, verify, and document the frontiers of knowledge.
As the cryptographic industry continues to mature, the transition from "conjectured security" to "provable security" will be the defining challenge of the decade. With the launch of better.codes, that transition has officially moved from a theoretical aspiration to a tangible, ongoing operation. For researchers, developers, and AI enthusiasts, the challenge is clear: the code is open, the target is set, and the frontier is waiting to be pushed forward.
