GonkaRouter integration

This page renders docs/GONKA-INTEGRATION.md from the repository at request time.

How OpenVerdict runs every piece of AI reasoning on the Gonka Network through the official gateway (gonkarouter.io), and how a judge can verify that claim independently in about two minutes.

The short version

Every juror inference in every verification runs on GonkaRouter. There is no other inference path in the codebase: no OpenAI, no Anthropic, no local model, no fallback provider. The jury is five seats drawn across three Gonka-hosted model families (DeepSeek, Kimi, MiniMax), each seat votes independently under commit-reveal, and every single model turn's gateway request id is recorded, sealed, revealed and shown in the UI.

Where the integration lives

What the adapter guarantees

• Model pinning. Every request sends X-Gonka-No-Fallback: true, so a saturated model returns a real 429 instead of silently substituting a different model. Our retry and hedging logic absorbs the 429; a substitution can never be recorded as the requested juror. If the gateway ever reports a fallback via X-Gonka-Fallback, that notice is captured into the run's audit record.
• Identity capture. For every completion the adapter records the gateway x-request-id, the x-devshard-id, the completion id (the devshard-<n>-<seq> shape, kept verbatim as the Gonka request id) and the node system_fingerprint. These land in the sealed run bundle and on the public run view, per turn, not just per run.
• Neutrality by prompt and by protocol. The system prompts require evidence for and against from primary sources with mandatory citations; the prompt text is hashed over canonical JSON and bound into each juror's on-chain manifest, so the exact instructions are provable and immutable. A verdict without a verifiable citation fails closed and is never counted.
• Models never fetch. All web research is executed by the engine (SSRF-guarded, size-capped) and recorded; the models only read recorded content. Salts and seal keys never leave the engine.

Multi-model consensus

Five seats across at least three model families are drawn with Sui native randomness. Each seat researches independently and commits a sealed blake2b-256 vote on-chain before any vote is revealed, so no model can see or anchor on another's answer. A 4-of-5 supermajority settles; a split jury triggers a discussion round and a second sealed vote; a still-split jury settles as UNRESOLVED rather than forcing an answer. The Truth Score (0 to 100) is the confidence-weighted aggregate of the revealed votes and is minted into the on-chain resolution certificate.

The table vote

When a jury splits in round one, it debates in public (up to three exchanges, stopping early once nobody moves) and then round two is a second kind of GonkaRouter call: one sealed chat completion per juror, no tools, over a pinned prompt (TABLE_VOTE_PROMPT_SPEC_V1, hashed into manifest v6 alongside the research prompt). The juror sees only what is already on the table (the round-one record and the debate transcript), casts one vote, and that run is sealed, committed, revealed and verified exactly like a research run, just without a research trail.

How to verify our Gonka usage yourself

1. Open any finalized claim and click a juror, then "Open proof". The run provenance shows the requested model, served model, devshard, system fingerprint and gateway request id; the research trail repeats them per turn.
2. Press "Check with GonkaRouter". The page cross-checks the recorded request id against GonkaRouter's public receipts endpoint (GET https://api.gonkarouter.io/v1/receipts/<x-request-id>, no auth) and prints the direct URL so you can bypass our server entirely. The receipt confirms model, devshard, timing and outcome for that exact request: proof the inference ran on the Gonka network, not on a centralized server we control.
3. Open the claim's full report for the on-chain record (claim, committee, certificate and revealed-vote objects on Suiscan; sealed and revealed bundles on Walrus), or use the Audit page to recompute every hash in your own browser.

Claim extraction (URL input)

Pasting a page URL into the verify bar calls POST /api/extract-claim: the engine fetches the page through its SSRF-guarded evidence fetcher, sends the text to a Gonka-hosted model at temperature 0, and returns the single most check-worthy claim as one falsifiable sentence together with the Gonka request id of the extraction call itself. The user confirms or edits the claim and submits it through the normal statement-only path; the jurors then research it from scratch, so the submitter's page never becomes privileged evidence.

Live example (2026-08-31, production): posting {"url":"https://simple.wikipedia.org/wiki/Bitcoin"} returned "In June 2021, El Salvador became the first country in the world to make Bitcoin a legal tender." with modelId deepseek-ai/DeepSeek-V4-Flash-0731, gonkaRequestId devshard-67806-387 and gatewayRequestId req-1788185908680461519-129576, both checkable against GonkaRouter's public receipts endpoint.