An ontology language for AI systems.
AI systems fail in three recurring ways: agents that behave differently on identical inputs, models that assert things that do not exist, and systems that cannot show an auditor why they did what they did. All three are failures of meaning.
Arkhe is a neutral, version-controlled language for declaring what exists in a domain, how it connects, and what may be done to it, by whom, with what trace. One YAML file in a git repo is the authoritative artifact. A compiler turns it into tool contracts, a neutral intermediate form, and from those generates the artifacts AI systems consume. The spec is a contract, never a runtime.
arkhelang.org: overview and, in time, the rendered spec
| Repo | What it is |
|---|---|
| arkhelang | The language: specification, ADRs, golden-file fixtures, and the reference implementation |
| arkhelang.org | The website (Cloudflare Pages) |
Packages: PyPI · npm · crates.io arkhelang / arkhe
Early development. v0.1 ships four things: the spec, a validator
(arkhe validate), the tool contracts, and one emitter. Design decisions
are public as ADRs.
Arkhe's author also authors GATE, an open framework for governed agent runtimes. Interoperability is a goal; exclusivity is not. Arkhe is runtime-neutral and GATE gets no privileged hooks.
Apache-2.0. Free forever. No CLA. A personal open-source project by Andrew Stevens.
