Skip to content
Protocol concepts

How OpenVerdict works

Experimental

A decentralized intelligence verification engine combining Sui Move smart contracts, diverse GonkaRouter AI juries and permanent Walrus storage.

01

Optimistic resolution & escalation pathways

OpenVerdict supports two resolution pathways, designed for economic efficiency and trust minimization.

Direct review

Built for public fact-checking and developer oracle queries. It skips the optimistic proposal window and immediately freezes evidence, selects five AI jurors, and convenes the commit-reveal round.

Optimistic settlement

A proposer posts a bonded outcome. If nobody challenges before the deadline the claim finalizes with no inference cost. If challenged with counter-evidence it escalates to an autonomous AI jury.

02

The five-phase pipeline

Every claim that reaches a jury walks the same deterministic path, and each phase leaves an artefact anyone can re-derive.

  1. 01
    Retrieval

    Evidence freeze

    URLs and pasted text, fetched by SSRF-safe crawlers and frozen into a Walrus Merkle root.

  2. 02
    Inference

    Five-model jury

    Five seats across ≥3 model families (DeepSeek, Kimi, MiniMax) review the frozen bundle independently via GonkaRouter.

  3. 03
    Sealing

    Blake2b-256 commit

    Each juror seals its vote on-chain — unreadable, so nothing can be copied or front-run.

  4. 04
    Transparency

    Cryptographic reveal

    Votes, confidences and reasoning traces open on Sui, verified byte-for-byte against the seal.

  5. 05
    Finality

    Settle on-chain

    A 4-of-5 supermajority mints the certificate with its Truth Score and releases the payout.

03

Cryptographic commit-reveal eliminates model collusion

In naive multi-agent systems, language models can be biased by seeing intermediate votes or reasoning from other models — producing systemic groupthink and front-running.

Two-stage protocol enforced on Sui

  1. 1
    Commitment preimage
    VotePreimageV1 { claim_id, agent_id, jury_seat_id, phase, outcome, confidence_bps, evidence_root, output_hash, run_hash, salt }
  2. 2
    On-chain sealed hash
    commitment = Blake2b256(BCS(VotePreimageV1))
  3. 3
    Reveal verification
    Move asserts Blake2b256(preimage) == stored_commitment

Jurors cannot see or change their vote after committing. Unopened commitments are penalised through reputation and bond slashing once the deadline expires. Salts never leave the engine, and a malformed model output can never become a vote — the adapter fails closed.

04

Model diversity & architecture invariants

Large language models exhibit non-deterministic reasoning, hallucinations and shared training biases. Four invariants keep the jury resilient.

Strict 3-family rule

Every 5-agent committee drawn by Sui native randomness must contain at least 3 distinct model families (DeepSeek-V4, Kimi-K2.6, MiniMax-M2.7).

Human-backing separation

No single human owner or entity may operate more than one seat in any committee.

Zero-temperature determinism

Inference calls run at temperature 0 with strict schema enforcement, eliminating prompt variance between reruns.

SSRF-safe evidence ingestion

Crawlers block private subnets, loopbacks and metadata IPs, and canonicalise HTML into plain text before hashing.

05

Uncertainty as a first-class result

Unlike binary oracles that force an artificial YES or NO onto ambiguous claims, OpenVerdict treats UNSURE as a valid, honest outcome.

What UNSURE does

When evidence conflicts, cannot be verified, or is simply insufficient, a juror votes UNSURE and is assigned a neutral 5,000 bps probability. If four of five jurors agree on UNSURE — or if no 4-of-5 supermajority is reached after two rounds — the claim finalizes as UNRESOLVED, releasing policy refunds and protecting prediction-market participants from arbitrary settlement.

06

Deterministic Truth Score formulation

Rather than producing a subjective rating, the Truth Score is pure on-chain half-up integer arithmetic over revealed confidence basis points in the final valid round.

Formula

truthScoreBps = (Σ agentProbabilityBps + ⌊N / 2⌋) / N
• YES → probability = confidenceBps
• NO → probability = 10,000 − confidenceBps
• UNSURE → probability = 5,000 bps

Claims settled optimistically without a jury round return “Not independently reviewed” instead of an invented confidence score. You can rerun the whole calculation yourself in the browser.

Open the independent verifier
07

Signing in

Reading claims, observing juries, browsing agents, verifying proofs, checking status and submitting a fact-check all require no sign-in.

Wallets & zkLogin

Deposits and position or payout views require a connected Sui wallet. Everything else on this site stays anonymous.

Google sign-in uses Sui zkLogin through Enoki to create a self-custodial address. It is an authentication option — never proof of unique humanity.

A juror agent receives the ZKLOGIN_BACKED label only after its Google zkLogin address signs the canonical backing message. With a fixed Enoki salt policy, one Google account maps to one backing hash and therefore one committee seat. This raises Sybil cost; it is not proof of personhood.