ĀML is currently stewarded by ĀRU Intelligence Inc. and remains a research prototype. Governance exists to move ĀML toward an independently implementable, testable standard rather than a single-vendor codebase.
- Specification and implementation are separate.
aml-coreis the reference implementation, not the definition of correctness by itself. - Compatibility claims must be testable. Published conformance fixtures, protocol vectors, and compatibility levels are the basis for interoperability claims.
- Breaking changes require versioning. Wire identifiers, schemas, registry entries, and accepted RFCs must not change silently.
- Independent implementations are encouraged. A useful independent runtime may expose ambiguity or flaws in the reference implementation.
- Dissent is evidence. Security failures, policy disagreements, accessibility findings, and interoperability mismatches should result in documented issues and regression tests.
- Integrity is not morality. Hashes and signatures prove integrity/key possession, not truth, ethics, legality, wellbeing, or accessibility.
- Human-impact claims require evidence. Present attention/restoration values are declared model inputs, not validated universal measurements of cognition or wellbeing.
Draft → Experimental → Candidate → Accepted → Superseded/Withdrawn
A Candidate RFC should have:
- a reference implementation;
- reproducible tests or protocol vectors;
- security considerations;
- compatibility and migration notes;
- at least one independent review.
An interoperability RFC should not become Accepted until at least one independent implementation or protocol peer demonstrates compatible behavior.
Stable protocol identifiers live in protocol/registry.json.
Changes to stable media-type candidates, wire versions, message kinds, or required conformance capabilities should:
- be proposed through an RFC or explicit registry-change issue;
- preserve existing identifiers unless a new version is introduced;
- include discovery/registry consistency tests;
- document backwards compatibility;
- avoid reusing an old identifier for new semantics.
High-impact future governance actions may use ĀML threshold-authorization primitives so multiple distinct keys are required to approve a registry or policy change. The repository does not claim that governance is already decentralized.
Compatibility levels are defined in conformance/levels.json and RFC 0008.
An implementation should identify the exact:
- AML language/spec version;
- wire version;
- conformance level;
- implementation/version;
- reproducible test artifact or test date.
Compatibility must not imply endorsement, regulatory approval, or certification of ethical behavior.
Security-relevant changes should include:
- threat model impact;
- replay/tamper considerations;
- canonicalization/signature impact;
- trust-root or delegation impact;
- protocol downgrade considerations;
- regression tests when practical.
Public adversarial review is encouraged. Sensitive vulnerabilities should use the repository security-reporting path rather than public disclosure before remediation.
aruintelligence/aml-core is the current reference implementation. Independent implementations MUST NOT be required to import its internal modules to claim compatibility.
Code licensing and claimed marks are separate. Any future “ĀML Compatible” mark should communicate only compliance with published technical requirements.
Created by Daniel Jacob Read IV and stewarded by ĀRU Intelligence Inc.™.