Base Sepolia · Testnet
GET STARTED
You can explore Clary Coin and use its wallet without installing research software. Producing and submitting mathematical proofs requires a separate local research environment.
Public research participation and downloads are not open. As of October 20, 2026, the contracts and fixed checker are deployed, but the public proof-to-reward-and-transfer rehearsal is incomplete. There is no CLARY sale, exchange market, or promised mainnet conversion.
TWO WAYS TO PARTICIPATE
Read the rules, inspect the public contracts, and use a compatible Ethereum wallet on Base Sepolia. Holding or sending CLARY does not require Lean, Codex, or the research package.
The ERC-20 token is deployed, but its observed supply is still zero. The wallet interface does not itself create a balance or a market.
Open the testnet wallet →The research package includes Lean, the proof exporter and local checker, a bounded research worker, submission tools, and an MCP connection. It runs on your computer and uses your own approved AI service.
The tested package currently supports macOS on Apple silicon. Windows and Linux research packages have not been validated.
FROM PROOF TO REWARD
Check the exact problem, accepted proof format, score direction, improvement threshold, fixed prize, and block deadlines. DFT rewards a proven theoretical rate; a practical benchmark does not replace the Lean proof.
Write a proof yourself or use the configured AI research worker. Lean compiles the proof, the exporter produces its supported artifact, and the local checker checks the complete new proof against the pinned common library. Failed candidates do not proceed to submission.
A Base transaction commits to the artifact, exact score, submitter, beneficiary, and private random opening. Keep the proof and opening private during the commit phase. The supported client waits until the commitment and cutoff are finalized under its explicit two-RPC policy before disclosure.
The client sends the artifact to the fixed Rust checker on the Internet Computer. Its replicated execution checks the proof and derives the score. The service signs the exact checked result with its threshold key. This is verification, not encryption of the mathematical proof.
Publish the exact proof bytes through the artifact registry and reveal the matching opening and signed result before the reveal deadline. The Base verifier authenticates the ICP signature and artifact identity; it does not execute Lean itself.
After the deadline, anyone can settle the round. The best eligible score wins; ties use canonical commitment order. Anyone may execute the winning claim, but it always issues CLARY to the beneficiary fixed earlier. That balance can then be transferred using the ordinary token interface.
The tested research runtime is approximately 3.3 GB compressed and expands to roughly 9.8 GB before additional working files. Measured proof workflows have used several GB of memory. A known DFT source took about five minutes to compile, export, and check in one installed-runtime trial. Researching a new result can take much longer or produce no eligible improvement.
A genuine known DFT artifact was checked and signed on the public ICP service in about 13.1 seconds. This is one observed result, not a promised response time. Publication transactions, chain finality, deadlines, and the participant's configured limits add further time.
The initial verification service draws from a finite operator-funded pool. Your AI subscription, computer use, and Base transaction gas are separate. Base Sepolia gas uses test ETH. Capacity may be exhausted; CLARY issuance does not automatically replenish the service. A future paid verification route would require separate published terms.
AUTOMATION WITH LIMITS
The implemented worker proposes Lean proofs, checks them, and can return structured errors for bounded repair. The outer controller independently checks a candidate and coordinates the sealed submission, verification, publication, and claim.
The complete integration has been tested with a fixed offline model response and genuine Lean verification. Live Codex research has not yet been tested in this release. These trials do not demonstrate autonomous mathematical discovery, and connecting MCP does not currently enable public participation.
The release design uses explicit limits on AI calls, elapsed time, transaction count, and fees. A separately unlocked local signer handles approved transactions. MCP and the research model do not receive the wallet key or its password. Stops, exhausted budgets, and uncertain transactions do not silently reset on restart.
The public release needs actual finalized proof-to-reward cycles and token transfers, reviewed live-model use, admission of the general participant client, and final package validation. The release also needs a finite sponsored operating plan and the reviewed hosting of its website, source snapshot, and downloads.
The initial rounds use already known proofs to test the real system. They must not be presented as new discoveries or as evidence that ordinary participants have earned rewards. There is no download button or registration queue while these gates remain incomplete.
See what is deployed, what has been demonstrated, and what remains →
Use the published Base Sepolia contract identity when adding CLARY to a wallet. Testnet tokens do not require buying real ETH. Never send a seed phrase, private key, or wallet password to a website, model, or project operator.
Successful proofs become public, and chain addresses and transactions are publicly visible. RPC, AI, and hosting providers may receive connection information under their own policies. Clary Coin does not promise anonymity or guaranteed rewards. Review the trust assumptions and technical white paper before committing resources.