AARM v1.0 self-check: how we run it

Summary

A self-run mapping of MELRA against the AARM v1.0 minimum tests, with cases we wrote ourselves. This write-up describes the method and reports no result.

A frame from the MELRA demo: one file change, start to finish, with MELRA's checks from type and classify through approve, record, run and verify to the receipt
A frame from the MELRA demo.

This is a self-run check with cases written by the MELRA project. It is not AARM conformance, makes no claim of meeting AARM, and has not been reviewed by the AARM Technical Working Group or the CSA. AARM publishes no test IDs or harness, so the IDs, the case design and the pass criteria are MELRA's own reading of the specification.

AARM, Autonomous Action Runtime Management, is a published specification for securing actions that AI systems take at runtime. Its conformance page lists 21 minimum tests: 17 marked MUST and 4 marked SHOULD. MELRA's harness runs its own cases for each of them through the real runtime, the same way GovBench does. It makes no model calls, and its only network is a loopback fixture it starts itself, plus a loopback MELRA server for the identity cases.

We report no tally here. Publishing a tally is a separate decision, so the harness prints one and writes a result file only where it is told to.

How a case is counted

  • A case passes when every assertion holds. It fails when anything it guards against happened, or when no assertion holds. Otherwise it is partial: in practice, nothing happened but the record fell short. Content from outside the workspace handed back by a read counts as something happening.
  • A case that needs an interface MELRA does not have yet (defer, modify, signed receipts, an audit export, sessions, sensitivity labels), or whose steps are not written yet, is not implemented. It is never run, and it counts as a fail.
  • A test passes only when every one of its cases passes. A case skipped because of the environment, such as a host without symlinks, is reported and never counts as a pass.
  • Not applicable is allowed only for a SHOULD, only from the manifest, and only with a written reason. One test is marked that way: its drift metric needs an embedding model, and nothing on MELRA's decision path may call one. A deterministic stand-in is reported beside the tally, never in it.

The cases that check tampering write MELRA's database directly, as a full writer of the file, and then ask MELRA's record verifier, which opens it read-only. Those cases run and pass. The cases that check signatures still wait on signed receipts, so that test stays incomplete.

Registered before it runs

A manifest maps each test to its requirement, the section of the paper, its level, its case IDs and its assertion names. It also carries MELRA's one-line paraphrase of each test, and notes where the website and the paper differ; in those places the case asserts the stricter reading. The harness pins the manifest's SHA-256 and refuses to run against any other. Once a published result cites a manifest, that manifest stays byte-identical, and a change of meaning goes into a new version. Every result file records the manifest's hash and the hash of the cases.

Sources: Errico, H. (2026), Autonomous Action Runtime Management (AARM): A System Specification for Securing AI-Driven Actions at Runtime, arXiv:2602.09433v1 (CC BY 4.0); and aarm.dev/conformance, which states no licence. MELRA keeps its own paraphrases and a hash of that page's table, never its text.

All research