Execution oracle for AI agents: "is it safe to send NOW?" 12 networks measured every 2 s against independent RPC nodes, signed answers, on-chain registry on Arbitrum, pay-per-call in USDC via x402 on Arbitrum One. ETHGlobal + Casper finalist.
PhoenixOracleRegistry on Arbitrum Sepolia
Live links: 1. Live product: https://rtt.phoenix-ai.work — signed live verdict for Arbitrum (no signup): https://rtt.phoenix-ai.work/api/v1/demo/safe?chain=arbitrum 2. On-chain registry (Arbitrum Sepolia): https://sepolia.arbiscan.io/address/0xD394fFAE51d8fB52187Cf3aE3b014Ddc80Dc7B15 3. Source code: https://github.com/kant19801201behax5/x402-health-oracle Price oracles answer "what is it worth?". Phoenix answers "should an agent execute right now?". A lagging or overloaded RPC never reports its own problem — the transaction just lands late, reverts or hangs. Phoenix measures 12 networks (Arbitrum, Base, Optimism, zkSync, Scroll, Linea, Mantle, Blast, Mode, Taiko, Polygon zkEVM, Casper) every 2 seconds against 2–3 independent RPC nodes per network (Casper: 10), counts only real JSON-RPC answers and returns a signed verdict: safe / reason, p95/p99 latency, divergence between nodes, node lag behind the head, request loss and transaction revert ratio. Three evidence layers, ~2 million records per day, each signed BLAKE3 + Ed25519: - Kernel: eBPF uprobes time every TLS round trip to every RPC node (kernel_tls_rtt in paid answers). - Network: per-node JSON-RPC latency, p95/p99, node divergence, head block height and age, halted-chain detection from the chain's own block timestamps (Polygon zkEVM is correctly reported as halted). - Execution: revert ratio from block receipts (Arbitrum, Base, Optimism). Paid answers are signed EIP-191 and verifiable offline. Latency agrees with an independent same-host measurement within 0–7 ms; revert ratios match an independent recount from block receipts. On Arbitrum: - PhoenixOracleRegistry on Arbitrum Sepolia exposes isSafe() / riskLabel() / the latest snapshot to smart contracts; the publisher writes the live verdict on-chain every 5 minutes. - The publisher runs inside a kernel sandbox (BPF LSM: no new executables, outbound connections only to allowed ports) — verified live: it keeps publishing under the guard. - Agents pay per call in USDC on Arbitrum One (x402 v2, eip155:42161), as well as Base and Polygon. Silicon DNA (built in): classifies every caller — HUMAN / LEGIT_AGENT / MALICIOUS_BOT / SYBIL — with a 14-layer behavioural analysis of request timing and headers, an ML-KEM-768 post-quantum session channel and a signed agent passport. Distribution: x402 API ($0.01 USDC per call, free delayed demo), MCP server (npm phoenix-mcp-server 1.3.0), OpenAPI 3.1, llms.txt, .well-known/x402. In the week of Sep 22–29 the paid endpoints received 17,028 requests from 380 distinct clients (x402 directories and monitoring agents). Infrastructure: Cloudflare → nginx → DigitalOcean NYC, 14 production services, internal ports on localhost only; every deployment verified by sha256 against git, checked live (including a forged-payment test) and rolled back automatically on failure. Paper: Zenodo DOI 10.5281/zenodo.22239862.