{"schema":"merit.cvo/v3","engine":"merit-jury/0.1.0","what":"A premium 4th verification tier ON TOP of the cheap cascade (numeric → NLI → single judge). Decompose an answer into atomic claims; only the borderline claims escalate to a panel of models from DIFFERENT families (0G DeepSeek/GLM/Qwen/MiniMax/Kimi, so errors are uncorrelated); a reputation-weighted supermajority decides each claim; settlement is GRADED per claim (pay the claims that pass, never all-or-nothing weakest-link).","roster":["deepseek-v4-flash","glm-5.2","qwen3.7-plus","minimax-m3","kimi-k2.7-code"],"certificate":"Returns a signed merit.cvo/v3 certificate: a Merkle root over every ballot, each ballot carrying its 0G attestation handle (the TEE-attested completion's on-chain provider address + request id + response key), so the panel is independently checkable.","economics":"Cost ≈ borderline-claims × jurors (open-model calls, 1–3s each). Use it for high-value settlements, not sub-cent metering.","settlement":"The jury produces the GRADED settlement DECISION (gradedUsdc per claim) — a signed certificate a rail/settlement-hook releases on. The jury itself moves no USDC (like /api/mandate/settle, it clears; it does not pay).","priorArt":"Composes Mira Network / Atoma consensus-inference (decompose → diverse panel → reputation-weighted consensus). Merit's novel part: binding that consensus to the on-chain payment-settlement decision (the signed certificate a rail/hook releases on) beneath a deterministic numeric $0 floor.","keyed":"POST requires Authorization: Bearer <key> — it convenes a paid model panel (real model cost) and produces a settlement decision.","forcePanel":"Set forcePanel:true for the high-assurance tier — every claim the numeric floor clears goes to the full panel, bypassing the free NLI shortcut (NLI still runs and is recorded).","post":"POST { source, answer | claim, amount?, question?, models?, threshold?, forcePanel? }"}