AMS-1 verification badges
A badge states one thing and one thing only: how far a verification artifact has traveled on the adoption ladder. There is no composite score, no A–F grade, and no “trust level” — safety, quality and adoption are different questions, and blending them is how rating systems mislead. Promotion happens only on evidence recorded in the artifact itself, never on payment, stars or age.
Stage 0 · self-attested 
The artifact exists and its inputs are hashed. The starting point — every verification begins here, including ours.
Stage 1 · recomputable 
Inputs hashed, rules version recorded: a third party with
the same inputs and the same registry reproduces every number. Generated by
agentmeasure verify, graded by agentmeasure certify.
Stage 2 · independently reproduced 
Someone else re-ran it and recorded the digest — the reproduction entry must match the artifact’s deterministic digest. Not self-declared: the digest math has to agree.
Stage 3 · counterparty accepted 
The other side of the bill accepted the evidence — a vendor confirmation, credit note or payment is recorded in the artifact. The rarest and the only stage that involves the counterparty.
How to get one
Run the verification, grade the artifact, embed the badge for the stage the evidence supports:
The grader names exactly what is missing for the next stage. Claiming a higher stage than the evidence is checkable by anyone — the digest is deterministic.
Who runs AgentMeasure vectors
A snapshot of external adopters — projects and maintainers who shipped AgentMeasure conformance vectors, fixtures or corrections (each links to the artifact):
- Urusilla (langfuse discussion) — two external conformance vectors, including the one that broke our validator and got it fixed
- iwasinnam2 · Metrecept (litellm #39057) — signed-receipt semantics from a production metering corpus; five limits encoded as vectors
- damian-112 · iLands — first full independent reproduction (236 tests, 14 vectors, pinned commit) and the AgentOps cache-drop independent audit
- razzlo · Rulestack — the $436K→$54K real-bill recount that became our first real-world evidence case
- Chandan Kumar · modelpath — latency-basis vector merged as external fixture 002
- David Turner — inference-boundary fixture; Rajesh Sharma (OWASP) and Gunjan Jaswal — DR-005 evidence-grade refinements
Snapshot of 2026-10. Ran a vector, found a discrepancy, shipped a fix? Open a pull request — the list is the artifacts, nothing else.