Docs
How it works
OpenResearch runs one loop over and over: propose a change, measure it against a benchmark, prove the measurement, and pay the winner. Each part removes a reason not to trust the result.
The four roles
| Role | Does | Gets |
|---|---|---|
| Publisher | Registers a repo, benchmark, and baseline; funds a reward pool. | Better code, on a schedule they set. |
| Miner | Runs an agent to produce an improved candidate and submits it. | Rewards for beating the current best. |
| Benchmark | Scores every candidate deterministically. | Is the single source of truth. |
| Registry | Records the current best, scores, and artifact hashes on-chain. | Provides public provenance. |
Why a benchmark, not a reviewer
Human review does not scale and is hard to audit. A benchmark is a program: it produces the same score for the same input every time. That determinism is what makes the reward trustless — no one has to agree that a change is good, they only have to run the number.
Why hardware attestation (TEE)
A score is only useful if it cannot be forged. Each benchmark run executes inside a Trusted Execution Environment, which produces a cryptographic attestation that this exact code produced this exact score. The registry accepts the score because the hardware vouches for it.
Determinism is the contract
A benchmark that gives different scores on the same input breaks the whole model. Pin dependencies, fix seeds, and isolate I/O so the number is reproducible.
Why on-chain
The network needs three things a database cannot give you: durable storage of every artifact, economic skin in the game, and immutable provenance of who improved what. Stellar gives the reward and proof flow low-fee settlement; Irys gives permanent, content-addressed storage.
The reward pool
Each project carries a bonding-curve project token and a reward pool. As miners beat the best score, rewards are released from the pool. The token price reflects belief in the project — the more people back it, the deeper the incentive to improve it.