AI statement

No model is trained, embedded, or called at run time. Nothing performs inference, and nothing in library code reaches a network at all. AI is used to build the software — in the same sense a compiler and a linter are used to build it. The EU AI Act imposes no obligation on this project, because it binds providers and deployers of AI systems and this workspace is not one; the disclosure is voluntary, and no standard is claimed as conformity.

The tooling is agentic AI coding assistance — Claude Code, by Anthropic — in sessions the maintainer directs and reviews. Crate code, tests, specifications, and documentation are ai-generated in the W3C AI Content Disclosure vocabulary’s sense: AI-generated with human prompting and review. What the crates do and do not do — scope, non-goals, what a release claims — is decided by the maintainer, with no AI involvement — deciding that a crate should release at all stays his alone, unconditionally. As of 2026-09-02, once he has scoped and named a release, an agent may judge whether it is ready against this project’s own stated criteria and run cargo publish for it; that is ai-assisted, not autonomous — the document’s term for AI-generated without meaningful human oversight. Nothing is autonomous: no commit is automated, no agent publishes as a standing job or on its own initiative, and nothing lands on its own authority.

One named human is accountable for every change, whatever tool produced the bytes. Commit trailers record tool participation; the author field of every commit records accountability, and a tool is never named as an author or signer of anything.

Every change, whoever or whatever wrote it, passes the same gates: the test suite, pedantic lints denied, formatting, documented public items, the minimum-Rust-version build, and the trademark checker. The specification is the control that catches a plausible-but-wrong implementation regardless of who wrote it: every rule is numbered, backed by a test, and a tool cannot quietly redefine what correct means. Tests and expectations are never weakened to make a build pass.

In this project, AI is prohibited from:

• committing or merging anything on its own authority, or signing anything;
• deciding whether to accept a contribution from someone else;
• deciding what the HL7® v2 standard means where it is silent, or what a release claims;
• deciding that a crate should release at all, or publishing as a standing job or on its own initiative outside a session the maintainer is directing — judging whether an already-scoped, maintainer-named release is ready, and running cargo publish for it, is the one exception, permitted as of 2026-09-02;
• weakening a test, an expectation, a spec rule, or a gate to make something pass;
• adding a dependency to any of the three crates;
• generating a sample or fixture by paraphrasing a real message — synthetic means constructed, not laundered.

And the limitations are stated, because a disclosure without them is marketing: the gates prove what they test, not correctness; review depth is one person’s; provenance uncertainty survives; and the whole document is a self-declaration that no third party has audited. The checkable artefacts — the specs, the tests, the commit trailers, the published benchmark method — are the counterweight: they can disagree with the document, and if they do, the document is wrong.

The canonical document is AI_STATEMENT.md at the repository root — versioned, with an effective date, a change log, and a machine-readable summary; this page is a summary of it. No patient data, no personally identifiable health information, and no customer data exists anywhere in this project, and therefore in any prompt.