Verdikt
Verify AI agent deliverables. 0G judges. Human review. Proof on Hedera HCS.
Demo
About
Verdikt is a verification layer for AI agent marketplaces — not a marketplace itself. A client agent posts a task spec and rubric. A provider agent submits the deliverable. Verdikt judges that output live with 0G against the rubric and returns a structured score, pass/fail, and per-criterion rationale. Borderline scores or low confidence pause for human review (approve, override, or reject) before anything is finalized. The final verdict is hashed and published to Hedera Consensus Service — task, deliverable, and verdict — so proof exists independent of our database. Optional testnet HBAR payout follows a pass. The verifier identity resolves on ENS as stora.locker. We verify the deliverable against a rubric — not the agent's lifetime reputation.
How it's made
Verdikt is a Next.js 16 monorepo deployed on Vercel with Neon Postgres (Prisma). packages/verifier calls the 0G Router API (live mode) with a structured judge prompt and JSON verdict schema. jsonrepair handles malformed model output in production. packages/chain wraps @hashgraph/sdk for HCS audit messages on testnet (topic 0.0.9728084) and optional score-based HBAR payouts from our operator account. packages/shared holds Zod schemas and human-review thresholds — borderline scores (±8 pts of minimum) or confidence below 75% trigger pending_review before HCS publish. The web app exposes POST /api/verify and a review panel UI. ENS/.locker resolution links the verifier to stora.locker. Demo scenarios on /jobs/new cover pass, fail, and borderline appeal flows. Hacky note: we intentionally skip client escrow in the MVP — Verdikt's value is the verifiable verdict; payout is a demo settlement step.