The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Atinamos Evidence listing page.
Public research, schemas and technical evidence for Atinamos independent verification of paid AI-agent and machine services.
Human-readable research: https://verify.atinamos.co.uk
Technical evidence: this repository
You can pay to be tested. You cannot pay to be trusted.
Atinamos produces scoped evidence. The buyer decides what that evidence means under its own policy.
Atinamos does not issue a universal trust score, certification, approval, safe/unsafe verdict or permanent provider rating.
Read the canonical public status first:
The main technical milestone is now the first unattended generic external Assurance execution from an already-frozen plan.
One internal production coordinator invocation handled the routine paid execution stages without an operator manually stepping x402 challenge acquisition, payment signing, paid dispatch, settlement observation, fulfilment evaluation or Verification publication.
Independent Base settlement:
Signed evidence:
Observed evidence layers:
The service returned the advertised kind of JSON deliverable. Atinamos did not infer semantic correctness or quality from that fulfilment.
The Keyronne run began at PLAN_FROZEN.
It proves:
It does not yet prove the complete public seller journey:
That seller product journey is now the main development lane.
Human seller productisation comes first. A later AI-seller API/MCP should use the same underlying Assurance run/order model rather than introducing another execution engine.
On 6 September 2026, Atinamos completed the first end-to-end external deterministic commissioning of the current Assurance Runner → Verification publication path against:
Two bounded x402 calls were made at 0.005 USDC each. Settlement was independently observed on Base, useful fulfilment returned, and the returned valid field matched independently derived ISO 13616 MOD-97 expectations for one valid and one invalid fixture.
That run was human-triggered commissioning. It remains an important earlier milestone and should not be retroactively described as unattended.
On 28 August 2026, an Atinamos-operated bounded buyer was given a task, maximum spend and buyer policy. It independently searched an external x402 marketplace, checked exact-route viability, queried published Atinamos evidence, applied its own policy, selected an external seller without the human naming the seller, made a bounded payment, consumed the result and validated the purchased output.
Final clean run:
This proves the controlled Atinamos buyer experiment. It does not establish widespread external-agent adoption of Atinamos evidence.
Also on 28 August 2026, Atinamos observed an unrelated external client purchase Atinamos JSON Repair for 0.005 USDC through x402 and receive successful HTTP fulfilment.
This establishes external commercial use of an Atinamos machine service. It does not establish that the purchaser was definitely an autonomous AI agent.
The repository deliberately retains successful and unsuccessful observations, including:
See EVIDENCE_INDEX.md for the current consolidated index.
Atinamos keeps these evidence layers separate:
Guardrails:
Endpoint and HTTP method are material evidence identity where applicable.
AI agents and developers can query selected published Atinamos evidence through the read-only Atinamos Evidence MCP:
The MCP exposes published evidence and buyer-supplied policy evaluation. It does not expose wallet access, payment signing, verification-triggering writes or a universal trust verdict.
Atinamos exposes a bounded read-only projection of machine-service discovery observations:
Market Search reports observed listings and provenance. A market listing is not verification evidence.
Current signed Assurance evidence can be read using:
Endpoint and HTTP method are material evidence identity.
Historical evidence should remain valid in the format in which it was issued rather than being silently rewritten.
This repository is the deliberately sanitised public technical publication layer for Atinamos Verification and Assurance research.
It may contain:
It does not contain:
GitHub is not where Atinamos Verification runs. It is where Atinamos publishes selected technical evidence and documentation for public scrutiny and machine consumption.
Copyright © 2026 MotionFil-AI.
See LICENSE.md for details.