agent-evidence-vocabulary

A closed, versioned vocabulary for claims about an agent's execution: what it attempted, what the system beneath it observed or refused, how directly and from what vantage the claim was obtained, and how much of a population the claim covers.

It's for anyone who writes a predicate, a verifier or an evidence format for agents and needs each term to mean one thing across vendors, including teams who want to map their own fields onto it.

Install the Python package for vocabulary v0.4.0: pip install agent-evidence-vocabulary==0.4.0.

Quick start

Pin the registry at a release and check its digest:

curl -fsSLO https://raw.githubusercontent.com/probityai/agent-evidence-vocabulary/v0.5.0/vocabulary.yaml
echo "6555b8c6e6c4fa6ebe3c59f6e95619ec383f93a0615036bed9da8a49f6302e09  vocabulary.yaml" | sha256sum -c

Then read a term:

pip install PyYAML==6.0.3
python3 -c 'import yaml; v = yaml.safe_load(open("vocabulary.yaml")); print(v["evidence_dimensions"]["observation_vantage"]["values"])'

It prints ['substrate', 'artifact']. Every term carries a definition, its allowed values and a lifecycle status.

To map your own system's fields onto the vocabulary, copy crosswalk/TEMPLATE.yaml and run python3 scripts/validate_crosswalks.py before opening a pull request.

Status

Release v0.5.0, tagged October 16, 2026. It adds ten proposed terms. The registry file is CC0-1.0, and the rest of the repository is Apache-2.0. Every term starts as proposed and is promoted only when an independent system emits it; the rules are in GOVERNANCE.md. The first crosswalk came from an independent implementation of the Adversarial Execution Evidence predicate, in pull request 2.

Documentation