Open Research

                          Open Research: 100xing AutoResearch (by Andrej Karpathy) On-Chain

Scientific progress is still limited by access: access to labs, compute, institutional credibility, and the small number of people with time to test new ideas. Meanwhile, many meaningful breakthroughs in software-based research are measurable: faster training loops, better kernels, lower memory use, improved numerical methods, stronger compression, or more efficient bioinformatics algorithms.

Open Research starts from a simple idea: if a benchmark can objectively measure whether code got better, then improving that code can become a form of proof of work.

Instead of asking a central authority to decide whether research progress is real, Open Research uses the benchmark as the oracle. A project creator starts with a real GitHub repository, defines an immutable experiment protocol, establishes a baseline score, and publishes the benchmark. From there, anyone can run an AI coding agent locally to search for improvements. If their code beats the current best score, they can submit it to the network.

The key is trust. Miners are not trusted to simply claim results. Validators re-run the benchmark in secure environments and attest whether the improvement is real. Accepted improvements update the project’s best known result and reward the contributor.

The result is a decentralized research network where anyone with compute and curiosity can contribute measurable improvements to open scientific code, earn rewards, and push the frontier forward without needing permission from a lab, company, or institution.

Open Research is built as a protocol plus a set of portable Agent Skills that plug into AI coding environments like Cursor, Claude Code, and Codex.

Project Creation The autoresearch-create skill helps a researcher turn an existing GitHub repository into a benchmarked research project. It guides an agent through:

• reading the codebase,
• deriving an experiment protocol,
• generating a structured protocol.json,
• rendering a human-readable program.md,
• running the project in a sandbox,
• establishing a deterministic baseline score,
• and optionally publishing the project on-chain.

This avoids vague research prompts. Every project starts from real code that runs, with a benchmark miners can actually compete against.

Mining The autoresearch-mine skill runs the competitive improvement loop. A miner pulls the current protocol and code, lets an AI agent propose and implement changes, runs the benchmark locally, and only keeps changes that improve the score. When a result beats the network frontier, the miner can submit a proposal with the code hash, benchmark claim, and stake.

Validation The autoresearch-validate skill lets verifier nodes independently resolve submitted artifacts, rerun the benchmark, apply deterministic checks, and approve or reject the proposal on-chain. This keeps settlement deterministic: LLMs help generate and explore ideas, but rewards depend on reproducible measurement.

Protocol Stack The project uses:

Agent Skills for creator, miner, and validator workflows. Docker-style sandboxing for baseline and benchmark execution. Python harnesses for trial running and reproducibility. Solidity smart contracts for project registry, proposal ledger, staking, verifier allowlists, and project tokens. 0G Galileo testnet for deployed registry contracts. 0G Storage for content-addressed protocol, benchmark, code, and baseline artifacts. TEE-based validation as the practical first verification path, with zk-style proof systems left open as a future path.

The design choice is the split between creativity and settlement: AI agents are free to explore aggressively, but the protocol only recognizes improvements that survive deterministic benchmarks and independent verification.