Axionorm

Agent policy as code. Axionorm is a YAML-first specification for agent authority and context boundaries, with an installable Python reference implementation and an OPA/Rego decision backend.

Status

v0.1.0 is an experimental reference implementation. It ships strict schemas, digest-bound context review, a real OPA evaluator, and enforcement integrated with Parabiont Protocol and PARALAX MCP.

It is not a complete semantic DLP product, a compliance certification, or a production security guarantee. Treat it as a working reference for how agent authority should be expressed and enforced.

What It Decides

Axionorm answers two questions before context or tools move:

1. Context: which reviewed items may leave the session, given the declared topics, kinds, labels, size limits, and content checks?
2. Tools: may this named tool run, with this file extension, this byte size, and a valid scope?

Both decisions go through OPA. Unknown tools, invalid reviews, duplicate IDs, oversized payloads, missing OPA, and unexpected evaluator output all fail closed.

How the Pieces Fit

• Candidate is a JSON document of items an agent proposes to carry.
• Review is an operator manifest that records the SHA-256 digest of each approved item.
• Policy is a strict YAML document with context and tool rules.
• OPA evaluates action=context and action=tool requests against the packaged Rego rules.

Editing an approved item changes its digest and invalidates the approval. The reviewer cannot approve a whole chat export by accident and expect the content to stay approved after edits.

Design Boundaries

• YAML describes authority. It does not execute installation stages.
• Pattern matching is defense in depth, not a semantic classifier. A reviewer can still approve mislabeled sensitive prose.
• OPA decides allow or deny per item. It does not remove text, infer private facts, or enforce anything inside a provider account.
• There is no permissive fallback when OPA times out or returns something unexpected.

Read Next

• Install and run the reference implementation.
• Policy schema for every field and limit.
• OPA backend for how decisions are evaluated.
• Worked example for a complete policy you can copy.

Source: github.com/amiencoy/axionorm