EU AI Act · Article 12 · Record-Keeping

Article 12 requires logs your own system can't rewrite. The standard for that doesn't exist yet. The cryptography does.

The EU AI Act's record-keeping obligation for high-risk AI systems now takes effect 2 December 2027 — a deferral from August 2026, formally adopted by the EU in June 2026, for a stated reason: the compliance tooling wasn't ready. There is still no finalized technical standard for Article 12 logging. Two drafts are in flight — prEN 18229-1 and ISO/IEC DIS 24970 — and neither is complete.

That vacuum is the most important fact in AI compliance right now, and most teams are drawing the wrong conclusion from it. The deferral is not a reprieve. It is a signal about what regulators concluded: the logging layer most organizations run today cannot produce evidence anyone else can trust.

The problem Article 12 actually names

A witness testifying about itself

Strip the regulatory language and Article 12 asks one question: when a regulator, auditor, or opposing counsel asks what your model decided and when — can you prove your answer wasn't edited after the fact?

A log your own infrastructure writes, stores, and serves cannot answer that question. Not because your team is dishonest — because the proof is structurally impossible. The party that controls the write path controls the history. Every SIEM export, every append-only database, every internally-signed audit trail shares the same flaw: the system being audited is attesting to its own honesty. In evidence terms, it is a witness testifying about itself.

Courts stopped accepting that before regulators did. Mobley v. Workday ended the "we just provide the tool" defense. In Lokken v. UnitedHealth, discovery reached model-governance records predating deployment. The demand converging across employment, credit, and healthcare AI litigation is the same one Article 12 encodes: an independent, tamper-evident record of what the model decided, when it was live, and that nobody quietly rewrote it.

The engineering translation

What "outside unilateral control" means cryptographically

The phrase has a precise engineering translation. A decision record is outside the operator's unilateral control when four properties hold.

Signed at decision time

Each attestation carries an ML-DSA-65 signature (NIST FIPS 204 — post-quantum, because these records must still verify in 2035 when a court asks). The message is canonicalized under RFC 8785 before signing, so the signature binds exact content, not a formatting accident.

Distributed signing authority

Attestations are signed by a 3-of-5 threshold federation across five independent regions with per-node keys. No single node — including ours — can unilaterally produce or alter a valid attestation.

Anchored to a public ledger the operator doesn't run

Every attestation lands on Hedera Consensus Service with a consensus timestamp the producing party cannot move. Rewriting history would require rewriting a public network's history.

Verifiable without trusting anyone — including us

Anyone can run npm install @rubric-protocol/verify, point it at an attestation, and get an independent cryptographic verdict: signature validity, trust-anchor match, ledger confirmation, quorum satisfaction. Zero configuration. No API call to Rubric. If Rubric disappeared tomorrow, every attestation ever anchored still verifies.

That last property is the one no internal logging stack can replicate, and it is the property Article 12's language is groping toward. The regulation defines the outcome. The cryptography is what the outcome requires.

The timing argument

Why building now beats waiting for the standard

Teams waiting for prEN 18229-1 or ISO/IEC 24970 to finalize are betting that retrofitting verifiable logging onto a production AI system in 2027 will be fast. It won't be. Decision provenance has a first-mile problem: a record signed at decision time is evidence; a record reconstructed later is testimony. Every decision your system makes between now and the standard's arrival is either independently attested or it is gone.

The draft standards define outcomes, not implementations. Infrastructure that produces signed, ledger-anchored, independently verifiable records today satisfies any plausible final form of those outcomes — and the organizations running it will walk into 2027 with a corpus of evidence instead of a migration project.

Verify one yourself

Don't take a page about tamper-evidence on faith — that would be the exact mistake this page is about. A real ML-DSA-65-signed, ledger-anchored attestation, verifiable in your own environment with the open npm package. No account. No sales call. The math either checks or it doesn't.

Verify a live attestation