EU AI ACT · ARTICLE 14 · ANNEX III 2 AUG 2026

Proof, not measurement.

Decision logs are testimony. Scores are opinion. Receipts are evidence.

Article 14 requires high-risk AI systems to be under effective human oversight — a person must be able to decline an output, override it, and stop the system. When a regulator, court, or insurer later asks the one question that matters — “show me that a specific authorized human approved this irreversible action before it ran” — most oversight tooling can’t answer it.

Download the one-pager (PDF)See a receipt enforced

What most “oversight” tooling actually produces

A decision log — self-reported and mutable; testimony, not evidence. An oversight score or dashboard — a heuristic rating of how attentive the reviewer probably was, over self-reported telemetry; an opinion aboutthe human, not a binding of the human to the act. A “human-in-the-loop” toggle — proof a step existed, not that a named person authorized this action. Having a human in the loop is not proof the human exercised authority over a specific irreversible act. Article 14 is about the latter.

The bright line: measurement vs. proof

Evidence of oversight must do two things a score cannot: bind a named human to an exact action before it ran, and verify without trust — confirmable offline, with a public key, without trusting the vendor or the log.

Oversight score / dashboard / logEMILIA authorization receipt

ArtifactA number or narrative about the reviewA cryptographic proof of the authorization itself

EvidenceStatistical / behavioral — an indicatorMathematical — Ed25519 over RFC 8785 canonical JSON

What it bindsHow attentive the reviewer probably wasA named human → an exact action, pre-execution

Outsider-verifiable?Only by trusting the issuerYes — offline, with a public key. No backend, no vendor trust

EnforcementAdvisory — scores after the factFail-closed — no receipt, no execution

Gameable?Yes — telemetry is self-reportedNo — a forged or altered authorization fails verification

A score tells you the oversight was probably real. A receipt proves the authorization happened — and refuses to let the action run without it. The binding is load-bearing; the score is, at most, an annotation on top of it.

What EMILIA provides for Article 14

The audit artifact a regulator can actually check

Deterministic verification with a public key; the receipt is the evidence, reproducible by anyone, years later.

Override / intervention you can prove

Approve, decline, or stop captured as signed, tamper-evident, per-action proof — not a log entry.

Enforcement, not exhortation

428 — no receipt, no execution. Oversight that can be bypassed isn’t oversight.

Two-person control where stakes demand it

Quorum receipts (a cryptographic two-person rule) + scoped delegation with verify-time constraint enforcement.

Built on accepted standards

Ed25519 (RFC 8032), JCS (RFC 8785); expresses as JWS (RFC 7515) / COSE; loggable to a transparency service (SCITT). Open IETF Internet-Draft, Apache-2.0.

The meaningful-human-control thesis RR-1 maintainer credential

Honest scope: a receipt is necessary, not sufficient. It proves a named human authorized the exact action; it does not prove the decision was wise, lawful, or fully informed — and a quality signal (e.g. a third-party judgment score) can sit inside a receipt as one more claim. EMILIA supplies the load-bearing, verifiable binding the rest of an oversight program builds on. Engineering & standards material, not legal advice.