CRE Market Guard: Vector Similarity + DON AI Consensus for Prediction Markets

This project automates prediction market listing with a two-stage AI gate and onchain auditability.
When a user submits a market request (question, sources, resolution criteria), the orchestrator first runs vector similarity screening to prevent duplicate or conflicting markets before expensive model verification.
If the request is semantically too close to an existing unresolved/open market, it is rejected early. If it is clean, it enters a strict FIFO queue and proceeds to distributed verification.
At verification time, the orchestrator fans out the same normalized request to four independent LLM worker nodes (GPT, Gemini, Grok, Claude). Each node evaluates with the same prompt template, returns a signed report, and includes execution receipt metadata. The orchestrator validates signatures, prompt/parameter hashes, and quorum rules (3-of-4 Byzantine-tolerant style).
If quorum is reached, the result is finalized and submitted onchain; if not, it fails with explicit reason codes (e.g., insufficient registered nodes, endpoint failure).
This design solves two core problems in decentralized prediction markets:

1. duplicate market spam and unnecessary inference cost, and
2. opaque/offchain moderation decisions.
3. By recording vector screening evidence and final verification bundle data onchain, the pipeline is auditable end-to-end.

Frontend: World App Mini App + Thirdweb wallet flow for request submission. Backend Orchestrator (Bun/TypeScript): request intake, strict FIFO queue, vector-screening integration, distributed verification fan-out, consensus validation, and onchain writes. Vector layer: embedding generation + Qdrant similarity search for dedup/conflict detection. LLM layer: 4 worker services and 4 verifier services (provider-specific), running in Railway. Consensus: quorum-based aggregation of signed node reports (3/4). Onchain layer: DonConsensusRegistrySkeleton contract on World Chain Sepolia (tested through Tenderly Virtual Testnet). Observability: Tenderly Explorer used to inspect vector screening transactions, consensus/finalization transactions, and status transitions. Chainlink CRE usage: CRE workflow simulation/deployment for confidential operator verification flow, and CRE-style confidential HTTP execution path in worker runtime.