Autonomous Buyer Proof #3: discovery, evidence, policy and payment

Research question

Can a bounded autonomous buyer receive only a goal, a budget and a risk policy, then independently discover machine services, use evidence to decide whether any are eligible, select one without a human naming the endpoint, make a live payment and validate the returned result?

The experiment also tested an important negative case: if no discovered service satisfies the buyer's policy, will the buyer refuse to spend rather than weaken the policy?

Task

The task was deliberately small and objectively measurable: repair the malformed JSON-like input {foo: 'bar', broken: true,} while preserving the supplied data.

The required semantic result was exactly {"foo":"bar","broken":true}. Keeping the target deterministic made it possible to distinguish successful fulfilment from a result that merely looked plausible.

Fixed buyer policy

If no candidate satisfied those conditions, no purchase was the correct policy outcome.

Experiment history mattered

The final successful purchase was not the first thing that happened. The public record retains failed and no-purchase stages because they shaped the system.

Final clean run

Fresh external discovery returned eight JSON-repair candidates and ranked Keyronne first. The exact Keyronne route passed the non-paying x402 preflight. Published evidence showed one paid successful fulfilment and zero failed paid fulfilments, satisfying the unchanged policy.

Other candidates were evaluated independently; more than one service could satisfy the evidence policy. The buyer selected the highest-ranked eligible candidate rather than having a human provide the seller.

Observed purchase

Inspect the settlement transaction ↗

What the buyer actually validated

The seller returned the repaired data inside its documented value response envelope. The buyer extracted that value and required an exact semantic object match against the pre-registered expected result.

The final public receipt recorded fulfilment_validation.valid = true and proof3_complete = true.

What the experiment supports

This bounded experiment supports that an Atinamos-operated buyer could independently discover external sellers, consult already-published evidence, apply buyer-owned procurement rules, select an eligible seller without human seller selection, make a bounded live x402 payment and objectively validate the purchased output.

It also demonstrates that evidence can legitimately produce a no-purchase decision. The policy was not weakened merely because the first discovered candidates were ineligible.

What it does not support

This does not establish permanent reliability for Keyronne or any other observed provider, nor a universal safe or unsafe verdict. It does not show that every autonomous buyer should use this policy.

Most importantly, the buyer was operated by Atinamos. This experiment is not evidence that unrelated external agents had independently adopted Atinamos or were purchasing Atinamos services in the wild.

Public artefacts

The public research repository contains both the narrative record and a deliberately sanitised machine-readable receipt. Wallet identity, recipient-account details, credentials, private infrastructure and unpublished verification mechanics are omitted.

Read the full public research note ↗

Open the sanitised experiment record ↗

Inspect the machine-readable public receipt ↗

Interpretation

The important step in Proof #3 is not simply that an automated client could send an x402 payment. The experiment connected external discovery, route identity, independent evidence and buyer-owned policy into the decision that preceded the payment.

That is the decision-layer problem Atinamos is interested in: an agent may be technically capable of paying many services, while still having reasons to buy none of them.