Independent evidence for autonomous machine commerce

Autonomous agents can discover services, hold spending authority and move real money. That does not answer a more important purchasing question: should the buyer have spent the money, and did the seller deliver what it claimed?

Atinamos Verification is an independent research and evidence project built from practical experiments in agent-to-agent commerce. We study the gap between what a service advertises, what its payment contract requests, what happens after payment and what evidence a future buyer can use before spending.

You can pay to be tested. You cannot pay to be trusted.

Atinamos produces evidence. Trust is the conclusion the buyer reaches from that evidence.

Research, Studies and Evidence

The public site separates three different things:

• Research — the questions, concepts and wider agent-commerce experiments;
• Studies — specific timestamped verification cases against named machine services;
• Evidence / Findings — consolidated observations and the machine-readable evidence behind them.

This distinction matters because a research hypothesis is not the same thing as a service test, and a readable case study is not the same thing as the underlying evidence record.

Three independent paid-service studies

On 22 August 2026 the Atinamos verifier exercised three independently owned machine services using real x402 payment paths. The outcomes were deliberately retained whether they passed or failed.

These are timestamped observations, not permanent provider ratings. One pass does not establish lasting trust and one failure does not establish that a service will always fail.

• Study 1 — x402Node JSON Repair
• Study 2 — x402.direct Service Directory Search
• Study 3 — x402engine Web Screenshot

Browse all Verification Studies.

Machine-readable evidence

Buyer agents can query published Atinamos evidence directly using the trust lookup API rather than scraping these pages:

GET /v1/trust?endpoint=<service-url>

The response reports observations such as settlement, fulfilment, output validation and failure classification. Where a public case study exists, the registry also links to it. It does not return a universal trusted/untrusted verdict. The buyer applies its own procurement and risk policy.

What we are measuring

Evidence may concern identity, endpoint existence, reachability, advertised capability, advertised versus requested price, payment contract, contract changes, settlement, execution, fulfilment, result validity, repeated operation, recency, provenance and contradictions between registries or marketplaces.

A low-risk buyer may need very little evidence. A higher-value buyer may require recent paid fulfilment evidence, repeated observations or fresh verification immediately before purchase.

Why this project exists

Our experiments have already shown several different layers of machine commerce:

• a service can be listed and discovered without proving it will fulfil;
• a payment can settle without proving the returned result satisfies the advertised output contract;
• HTTP 200 can coexist with an objectively invalid primary artefact;
• a machine-readable description can materially affect autonomous selection;
• a platform can support agent-to-agent communication without supporting autonomous paid procurement;
• external evidence can influence which service a machine buyer considers.

The goal is not to create another five-star review site or a universal 0–100 trust score. We are building a public evidence base that developers, buyer agents, marketplaces and researchers can inspect directly.

Start here

• AI Agent Trust & Verification Research
• Verification Studies
• Evidence and findings
• Methodology
• What Does Trust Mean in an AI Agent Marketplace?
• What Is Machine Contract Optimisation (MCO)?
• Can One AI Agent Autonomously Discover, Pay and Buy From Another?
• What Should an AI Agent Check Before Spending USDC?

Current boundary

Atinamos Verification can now independently exercise external x402 services, apply bounded payment policies, observe settlement and fulfilment, validate structured and binary-output assertions, retain failures and publish machine-readable evidence.

What remains unproven is the strongest commercial claim: whether this independent evidence materially changes an autonomous buyer's procurement decision. The next decisive experiment is therefore a Decision Delta test: candidate services → Atinamos evidence → buyer policy → choose/reject → payment.

When Atinamos has not observed a service, our conclusion is simply: Atinamos holds no evidence. That is not a claim that the service is unsafe, fraudulent or untrustworthy.