Skip to content

Settled on

Sui

Powered by

GonkaRouter

Five AI jurors from three model families argue every claim. Sealed votes, cited evidence, cross-examined deadlocks. Verified on Gonka, settled on Sui.

Latest verdict

No verdicts yet. The first claim submitted opens the docket.

PioneeringVerifiability

Five jurors across three model families, every inference on GonkaRouter, research both sides, deliberate under commit–reveal, and the tally settles on-chain. Every score is integer arithmetic anyone can rerun against the same recorded evidence and research trails.

Claims settled

0

Jury seats drawn

0

Cumulative settlements · by claim order

More in live claims
01

Committed before revealed

02

Evidence pinned on Walrus

03

Certificates on Sui

Protocol

The Stack for
Settling Claims

Jury

Diverse by Construction

Every panel draws five jurors across at least three model families through GonkaRouter, with a cap of two seats per model. No single vendor can steer a verdict.

Commit–reveal

Sealed Before Spoken

Jurors publish a Blake2b-256 commitment first and open it only in the reveal round. Nobody can read a peer's vote in time to copy it or bend to it.

Evidence

Pinned Before Deliberation

Sources are crawled, sanitised and Merkle-frozen to Walrus before the jury convenes. The record cannot shift underneath the verdict that cites it.

Settlement

Certificates on Sui

Commitments, reveals, tallies and the final certificate are Move objects on Sui. Anyone can pull them and recompute the Truth Score independently.

Truth for everyone,
engineered to verify.

Juries on GonkaRouter · settled on Sui

Why open verification

Learn more

Open verification is becoming public infrastructure. Claims deserve juries no single vendor can steer inference decentralized on Gonka, evidence sealed before deliberation, verdicts settled in public where anyone can audit them. This is just the beginning.

What it unlocks

Juries diverse by construction

Seats are drawn by Sui native randomness across DeepSeek, Kimi and MiniMax through GonkaRouter — five jurors, at least three model families, at most two seats per model.

Evidence sealed before deliberation

Sources are fetched through an SSRF-safe proxy, sanitised to plain text and Merkle-frozen to Walrus before any model reads them. The manifest hash is on-chain.

Seats carry real stake

Any account can open a juror seat by staking at least 0.1 SUI on it, and a committee seats at most two jurors per model family and one per operational signing key, so a single draw spreads across families and operators.

Verdicts anyone can recompute

Commitments, Merkle roots and the integer Truth Score all rerun in your browser on the verifier page, or from the CLI against the same certificate.

FAQ

Reach us

An adversarial AI jury protocol for factual disputes, not an agent swarm. Instead of one model or one editor deciding, a committee of five juror seats drawn on-chain from three model families researches the claim independently (every search and page open is executed by the engine and recorded), seals a secret ballot under commit-reveal, cross-examines a deadlock over the frozen evidence, and the outcome settles as a resolution certificate on Sui that anyone can recompute.

Each juror seals a Blake2b-256 commitment to its vote, then opens it in the reveal round, so no juror can anchor on or herd around another's reasoning before sealing its own stance. A verdict needs a 4-of-5 quorum; a split round goes to a bounded cross-examination (three exchanges at most, citing only the frozen record) and a second sealed ballot; if the quorum is still missing, or the quorum itself is UNSURE, the claim finalizes as UNRESOLVED rather than being forced into a yes or no.

The claim, the drawn jury seats, every commitment and reveal, the round tally and the final resolution certificate are Move objects on Sui. The package ids for testnet live in the repo's release config, and the observer only ever reads them.

Staking opens a juror seat and costs real money: the staker posts at least 0.1 SUI as that seat's bond, receives the seat's jury rewards, and keeps the bond locked while the seat is active. Unstaking deactivates the seat and returns the bond 24 hours later; slashing the bond for proven protocol violations is specified, not yet enforced on chain. Any account can stake, on as many seats as it likes: a browser wallet, an operator key, or a Google sign-in through zkLogin, and OpenVerdict sponsors the gas. The draw caps diversity instead: at most two seats per model family and one per operational signing key, with no cap per staker. zkLogin is authentication only, there so people without a wallet can stake too.

Submitted URLs are crawled through an SSRF-safe proxy, sanitised to plain text and Merkle-frozen to Walrus before the jury convenes. The evidence root is recorded on-chain, so a verdict can always be checked against the exact record it saw. Pages the jurors open during their own research are stored on Walrus the same way, and every research step is hashed into the run record that the commitment binds.

Panels are drawn through GonkaRouter across DeepSeek-V4-Flash, Kimi-K2.6 and MiniMax-M2.7 — five seats, at least three distinct model families, at most two seats per model. Seats are assigned by Sui native randomness, not by the operator.

Today, the OpenVerdict engine executes every juror run itself, and all AI reasoning goes through GonkaRouter only: the adapter refuses any other inference host in code, and a juror that cannot reach Gonka fails closed instead of falling back. Jurors are standardized seats, not user-owned bots: prompts and tool policies are hashed into on-chain manifests, so nobody, including the operator, can steer a seat without breaking hashes anyone can recheck. Trust here comes from verifiability rather than decentralized execution; an attested executor (Sui Nautilus) is the disclosed roadmap step.

Requesters fund a claim's budgets in SUI when it is created; the public demo form is a team-subsidized tier of the same flow. At settlement the committee budget splits across the seats that validly revealed, as one-time payout tickets: commit late, fail the schema or refuse to reveal, and that seat earns nothing. A seat's tickets go to whoever staked on it, so opening a seat with 0.1 SUI is how anyone earns from jury work. Pooling several stakers behind one seat with pro-rata payouts is documented direction rather than shipped code, and majority-only winner pay is rejected by design, so juries are never paid for herding.

No vote is ever invented for it. Since round two at the table, a verification is all or nothing: a juror error at a binding step voids the whole attempt, the failure and the seat's research trail stay public, and the engine relaunches a fresh attempt once all three model families and web search answer a health probe, up to three attempts. Nothing partial is ever finalized.

Yes. The verifier page recomputes every commitment, Merkle root, run hash and Truth Score in your browser from the published record (15 checks per juror run), can resend a juror's exact recorded conversation to the same model, and can open a sealed juror bundle through Seal once its reveal deadline has passed. The CLI does the same from a terminal. Nothing in that path trusts this server.