A verifiable risk brain for AI agents: 22 pay-per-call services over x402 on X Layer + Base. Every answer ships a proof you can re-run; crash-validated 14.3x out-of-sample on 79,386 real accounts.
Quiver is a verifiable risk brain for AI agents: 22 pay-per-call services over x402, live on X Layer (USDT0) and Base (USDC). Agents pay for answers in-band, and Quiver's bar is that every answer must be checkable, not trusted. Deterministic services ship a proof envelope: echoed inputs, the engine codeHash, ground-truth self-checks, and a contentHash you can recompute byte for byte, secp256k1-signed. Live-data services ship a signed, timestamped observation envelope. When a signal can't be grounded, the service says DATA_UNAVAILABLE for free instead of faking a number, and the billing contract is enforced in code: no delivered answer, no settlement.
The desk spans perp liquidation and portfolio crash-stress gates, Kelly position sizing, arbitrage-free options analytics (Breeden-Litzenberger risk-neutral distribution, greeks, dealer gamma, variance risk premium), transaction and EIP-712 signature safety, live DEX microstructure (Kyle's lambda, Amihud, VPIN), prediction-market execution, DeFi protocol risk, a macro-event radar, wallet forensics, agent-controlled charts, and cursor-based wallet diffs, all from one endpoint.
The risk engine is validated against the real market, pre-registered and out-of-sample: we replayed 79,386 real accounts through the Oct 10, 2025 crash, locked thresholds, then tested two crashes the model never saw. Flagged accounts got liquidated 14.3x and 13.3x more often than cleared ones. Registered on-chain as ERC-8004 agent #5152 on X Layer, with real third-party paid usage, a public buyer review ("Excellent x402 delivery... a clear ELEVATED risk verdict"), and 242 automated tests. Full methodology: https://quiver-production-c3a8.up.railway.app/paper OR https://drive.google.com/file/d/1K44jmBBLyFed1qF6Ib62YRh8J-jMQ-xW/view?usp=sharing
We shipped v1 (10 services) before the original July 17 deadline. When OKX extended it, we treated the extra days as a full second build-and-QA cycle:
Grew the desk from 10 to 22 services, adding the risk-gate layer (perp-gate, portfolio-gate, size-gate, exec-verify, treasury and LP risk, on-chain attestation) on top of the analytics layer, each with ground-truth self-checks.
Validated the engine against the real market with hard anti-overfitting discipline: hypotheses and thresholds written to an append-only log BEFORE computing, a temporal cutoff, then a population-scale replay of 79,386 real accounts through the Oct 10, 2025 crash plus two out-of-sample crashes (Feb 2026, Jun 2026). Result: 14.3x / 13.3x relative liquidation risk for flagged vs cleared accounts.
Proved the payment layer twice: dual-rail x402 settlement (X Layer USDT0 via the OKX facilitator, Base USDC via CDP), all 22 services field-tested on both rails, every receipt on-chain.
Ran a 3-day REAL buyer desk against our own listing through the marketplace funnel: 900+ envelope verifications with zero contentHash mismatches and byte-identical reproduction of deterministic results. The buyer's audit surfaced a real settlement leak (facilitator success:true with no transaction hash, roughly 7% of revenue never landing on-chain); we confirmed it on-chain, fixed it, and live-verified the fix within hours. 242 automated tests green.
Shipped the reliability layer: keep-warm infrastructure, per-settlement observability logging, and a billing contract where invalid input and unavailable data are free by construction.
Not fundraising. Self-funded. The 22 services are live with real pay-per-call revenue on two rails (X Layer USDT0 and Base USDC), including genuine third-party buyers.