Atinamos Verification
Core Principle You can pay to be tested. You cannot pay to be trusted.
Explore verification

Proof #3 — Discovery Search Criteria

Why this is published

The discovery phase is only meaningful if the marketplace search is seller-neutral and reproducible. This page records the criteria used by the Proof #3 buyer before Atinamos evidence or buyer-policy filtering was applied.

The discovery implementation is in the public Atinamos-Agent feature/proof3-autonomous-buyer branch at:

agent/aws_bazaar_discovery.py

Marketplace searched

The buyer searched the Coinbase/x402 Bazaar through the existing AWS AgentCore MCP gateway.

The search explicitly required:

curatedOnly=false

so both curated and non-curated Bazaar resources were eligible for discovery.

Human-supplied task

The human supplied a goal, spend bound and risk policy. No provider or seller endpoint was named.

The discovery task was:

Repair the following malformed JSON while preserving its data:
{foo: 'bar', broken: true,}

The discovery phase was instructed to search for paid machine services capable of:

repairing malformed JSON
or
validating-and-repairing malformed JSON

Validation-only services without a repair capability were therefore not suitable for the task.

Explicit neutrality constraints

Before marketplace search, the discovery agent was instructed:

  • do not purchase or invoke any paid service during discovery;
  • do not favour any provider or brand;
  • do not use Atinamos evidence or any external reputation layer during discovery;
  • do not invent endpoint details that are absent from marketplace metadata;
  • omit candidates whose invocation contract cannot be determined from marketplace metadata.

This separation is deliberate:

external marketplace discovery
-> marketplace relevance order
-> Atinamos evidence lookup
-> buyer-owned policy

Atinamos does not create the initial marketplace ranking.

Ranking criteria

The discovery agent was instructed to rank candidates only by:

  1. Marketplace task fit — whether the advertised service actually matched JSON repair / validate-and-repair.
  2. Machine-contract clarity — whether an executable endpoint, HTTP method and request contract could be determined from marketplace metadata.
  3. Price — the advertised service price in the marketplace result.

The prompt did not include Atinamos trust evidence, seller preference, ownership, commercial relationship or a pre-selected service.

Candidate-output requirements

The discovery agent could return at most eight candidates. For each candidate it had to provide:

provider
service
canonical HTTPS endpoint/route
GET or POST
advertised price in USDC
executable request URL
JSON request body when applicable
brief reason for marketplace ranking

The invocation details had to be executable for the supplied malformed JSON and derived from marketplace metadata rather than invented by the buyer.

Observed result on 28 August 2026

The successful dry-run returned eight external JSON-repair candidates. The highest-ranked service was Keyronne JSON Repair at 0.001 USDC.

Atinamos itself did not appear in this JSON-repair candidate set.

Only after this external ranking was produced did the buyer query Atinamos Verification. None of the eight candidates had qualifying published Atinamos evidence, so the unchanged buyer policy correctly produced an autonomous no-purchase decision.

What this methodology does not prove

This search is one observed marketplace search under one task and one model/runtime configuration. It does not establish a permanent or universal Bazaar ranking.

Marketplace contents, service metadata, prices, usage signals and model behaviour can change. Repeated searches may therefore return different rankings even when the high-level criteria remain the same.

The purpose of publishing the criteria is not to claim ranking determinism. It is to make clear what the buyer was allowed to consider and, equally importantly, what it was forbidden to consider.