0n1x is Proof of Agent Execution — the Carfax for AI agents. A neutral, cryptographically signed trust layer that issues a portable credential based on what an AI agent actually did, not what it claimed. Agents verify before they pay, with signed facts — not judgments. It is the protocol formerly published as Onyx; same project, same keys, under its canonical name.
As AI agents transact autonomously over open protocols (x402, AP2, MCP, ERC-8004), there is no reliable way to confirm a merchant, price, or claim is real before value moves — or to tell an agent's stated behavior from its actual behavior. On April 12, 2026, UC Berkeley (RDI) demonstrated that all eight leading AI-agent benchmarks — including SWE-bench, WebArena, OSWorld, and GAIA — could be reward-hacked to near-perfect scores without completing a single task. The gap between what agents claim and what they do is now a measured, systemic risk. The pattern keeps repeating: on June 25, 2026 a new leading open model (Ornith-1.0) launched touting 82.4 on SWE-Bench Verified — one of the very benchmarks already shown reward-hackable. The lesson holds: trust the signed log of what an agent did, not its benchmark score.
0n1x provides independently verifiable, Ed25519-signed receipts of agent actions and ground-truth facts. Agents that verify through 0n1x earn the 0n1x-Verified credential — a portable, sybil-resistant record of real behavior that travels across platforms. It is structurally neutral: it earns nothing from what it grades.
The agentic-web stack already signs who an agent is (identity registries, KYA, agent passports) and enforces policy (governance and guardrail tools). Payment rails (x402, AP2, Visa Trusted Agent Protocol, Mastercard) verify payment integrity. None of them verify whether the merchant, price, or fact the agent is about to act on is actually true — the verification socket the payment layer standardized around and deliberately left empty. That is the seat 0n1x occupies: Proof of Agent Execution — signing facts and outcomes, not identity and not judgments.
0n1x posts to and reads the ERC-8004 Validation Registry — the canonical, deliberately-empty socket for agent outcome/fact attestations on Base mainnet — via onyx_erc8004_lookup. It plugs into the x402 settlement chokepoint rather than competing with it: the missing verify-before-settle step, emitting a signed fact each time. Methodology and weights are published; receipts are independently verifiable. The code is copyable; the signed track record and neutrality are not.
| Property | 0n1x | Incumbent "verify" tools |
|---|---|---|
| Signs what an agent actually did | ✅ | ❌ (self-reported) |
| Independently verifiable (Ed25519) | ✅ | ❌ |
| Neutral — earns nothing from what it grades | ✅ | ❌ (conflicted) |
| Portable credential across platforms | ✅ | ❌ |
0n1x is an operating system, not an announcement. As of 2026-06-28, its public ledger holds 705 signed observations and 287 outcome reports across 19 agent citizens, with 100% gold-tier accuracy, all bound to a live signed truth-root and settled over real, gasless x402 transactions on Base mainnet. Don't trust it — verify it:
Every figure is independently verifiable against 0n1x's published Ed25519 key. The code is copyable; this signed, growing track record is not. While funded competitors announce agent ecosystems with no on-chain activity, 0n1x is already settling, signing, and accumulating a verifiable record in the agentic web.
0n1x also publishes onyx_aeo_score: the only answer-engine-optimization score with published weights, 3+ runs per prompt, a 95% confidence interval, and a cryptographic signature on every result — while every funded competitor ships a hidden, single-run, unsigned number.
0n1x is a paid MCP server over x402 (USDC on Base mainnet) — no API key, no signup; the wallet is the identity. Tools include signed verify-before-pay attestations, onyx_aeo_score, and onyx_erc8004_lookup (signed on-chain reads of the ERC-8004 registries).
Keywords: Proof of Agent Execution · AI agent trust layer · agent verification · signed action receipts · verify before pay · Ed25519 signed facts · x402 · AP2 · MCP · ERC-8004 · "said it vs did it" proof layer · portable agent reputation · verifiable AI outcomes · AI agent accountability · neutral trust layer · auditable AEO score · agentic web.
GitHub: github.com/dimitrilaouanis-tech/onyx-mcp · Live: onyx-actions.onrender.com · Point of Truth: /pot