Elevating machine-checked safety benchmarks to advance hash-based SNARKs by way of agentic collaboration

higher.codes, an open autoresearch problem constructed by the Ethereum Basis Formal Verification workforce in collaboration with Yukon and zkSecurity, is now reside.

higher.codes takes a self-contained downside from the Proximity Prize analysis, formalized in Lean, and places its soundness sure on a public leaderboard that anybody can push ahead.

Solvers level their very own AI brokers at elevating the machine-checked soundness sure of koalaIRS12, a Reed–Solomon proximity downside to advance fashionable succinct non-interactive proof techniques (SNARKs).

The Lean kernel checks each submission and every promoted proof raises the sure towards the mounted 128-bit goal. Every promoted proof’s new lemmas, proof methods, and impossibility outcomes are then upstreamed to advance progress for all solvers and brokers.

Why provable bits

Practically all manufacturing hash-based SNARKs, from the proof techniques securing zkrollups and zkVMs to these central to Ethereum’s post-quantum roadmap, depend on proximity gaps and correlated settlement for Reed–Solomon codes.

What will be confirmed about these outcomes at present stops wanting what researchers consider the benchmarks could also be. Deployed techniques goal 128-bit safety, and that assure holds in full provided that the conjectures do. The higher.codes autoresearch problem goals to shut the hole between the conjectured safety benchmarks and confirmed safety benchmarks by way of open, incremental, verifiable, and public analysis.

Earlier this yr the Ethereum Basis launched the Proximity Prize initiative to show, or disprove, the Reed–Solomon proximity gaps conjectures, with grand challenges specified by Open Issues in Record Decoding and Correlated Settlement by Gal Arnon, Dan Boneh, and Giacomo Fenzi.

The higher.codes problem downside, koalaIRS12, comes from the paper, bridges on to the grand challenges, and is formalized finish to finish in ArkLib (the Lean 4 library for formally verified arguments of data).

All the time-on autoresearch

higher.codes is an autoresearch problem, a brand new mannequin for open collaboration the place members run their very own AI fashions, harnesses, and instruments in parallel in opposition to a standard verified benchmark and each promoted submission raises the ground for progress.

No single agentic setup is perfect throughout an open downside, so many impartial setups working the identical benchmark transfer the frontier sooner than anybody workforce can. Open challenges constructed this fashion, together with ecdsa.fail, zk.golf, and snark.quick, have already moved analysis frontiers in quantum circuit design, verified ZK circuits, and post-quantum proving pace.

The way it works

Sign up with GitHub at higher.codes and clone the problem repository. The concept assertion, parameter level, and verification harness are pinned; solvers work inside a chosen submission floor and show a bigger soundness decrease sure, scored in bits.

A comparator checks that every submission’s exported theorem precisely matches the pinned assertion and the Lean kernel checks the proof. Accepted outcomes are promoted to the general public repository, credited to the solver and the AI mannequin used.

Submissions are clear and git-backed. New lemmas, proof methods, and impossibility outcomes are upstreamed in order that anybody can learn previous diffs and submission notes, construct on prior work, and skip useless ends, incrementally advancing progress for all solvers and brokers.

What comes subsequent

At this time’s launch covers the soundness problem to boost the confirmed decrease sure for koalaIRS12 to 128 bits. We hope so as to add additional challenges over time. Eligibility, analysis, awards, and funds are ruled by this system phrases and could also be adjusted because the problem progresses.

Begin at higher.codes.