VAC Protocol & Athena · regulatory alignment map
The EU AI Act asks for provable human oversight and complete records. That is what this infrastructure produces.
A plain-language map of where the Verifiable Authority Chain and Athena's calibration layer are designed to support the Act's obligations — for deployers, providers, and the assessors who evaluate them.
Where the clock actually stands (post-"Digital Omnibus", formally approved June 2026): the Act entered into force in August 2024 and applies in stages. The heavy high-risk regime — risk management, logging, provable human oversight, conformity assessment — was due 2 August 2026 but has been deferred to 2 December 2027 for stand-alone high-risk systems (justice, essential services, biometrics) and August 2028 for AI embedded in regulated products. Two things did not move: the Article 50 transparency duties still apply from 2 August 2026, and the reason given for the delay is that the compliance ecosystem — standards, evidence tooling — was not ready. Building the evidence layer is precisely the work regulators just gave organisations sixteen more months to do; every major law firm's guidance says the same thing: use the time, do not wait for it.
The numbers that moved this onto every board agenda: the Act's penalties are set against worldwide annual turnover, whichever is higher — up to €35M or 7% for prohibited practices; up to €15M or 3% for non-compliance with the Act's other obligations, including the human-oversight, record-keeping and transparency duties mapped on this page; up to €7.5M or 1% for supplying incorrect or misleading information to authorities. And the exposure is not a 2027 story: the Article 50 transparency duties carry that 3% / €15M tier from 2 August 2026 — days away — while the deferred high-risk regime lands on the same penalty scale in December 2027. For a global enterprise, 3% of turnover is a number that converts "oversight we can describe" into "oversight we can prove" as a budget line. That is why the compliance market is suddenly loud about audit trails — and why the harder question underneath them, which authorised person oversaw this and can you prove it independently, is the one this infrastructure answers.
The third door. Left to the obvious options, organisations face a false binary: treat the 3% as a cost of doing business, or quietly shrink their AI use to shrink the risk — less intelligence to stay safe. Fine-grain governance opens the third door:
keep, or grow, the AI intensity — with per-user, per-decision control, oversight that scales with the risk of the work, and evidence a regulator verifies independently. Not less intelligence. Governed intelligence.
Walk the retrofit console — configure a deployment in five steps →
Human oversightArticle 14
High-risk systems must be designed so natural persons can effectively oversee them. The VAC seal is oversight made evidentiary: a biometrically verified, present person authorises the consequential action, at a ceremony weight proportional to its stakes, producing an independently verifiable record that the oversight actually occurred — not a policy document asserting that it should.
Record-keeping & logsArticles 12 & 19
High-risk systems must log events automatically and providers must retain them. Every VAC authorisation emits a signed Authority Action Record — who, what, under which conditions, when — lodged to an audit chain any party can verify without trusting the operator. Logging that is cryptographically checkable rather than merely stored.
Deployer obligationsArticle 26
Deployers must use systems per instructions, assign competent human oversight, and monitor operation. Bounded delegation is the mechanism: each agent's authority is scoped — purpose, amount, counterparty, time — under a named, verified person, so "who was responsible" has a cryptographic answer, not an organisational shrug.
TransparencyArticles 13 & 50
Systems must be operable with appropriate transparency, and people informed when interacting with AI. The authority chain makes the human-vs-agent provenance of any action inspectable: an action either traces to a verified human seal or to a delegated agent under one — and the record shows which.
Accuracy & robustnessArticle 15 · FRIA Article 27
High-risk systems must achieve appropriate accuracy and robustness; public-interest deployers must assess fundamental-rights impact. Athena's known-answer calibration provides measured, per-domain evidence of model performance — 10,000+ scored challenges across 32 domains — the kind of empirical basis accuracy claims and impact assessments are supposed to rest on.
What does apply on 2 August 2026Article 50 · and the question underneath
The deferral moved the heavy regime, not the direction — and Article 50's transparency duties still land on 2 August 2026: people must know when they are interacting with AI, and synthetic content must be disclosed. Meanwhile the market is racing to generate audit trails of what AI systems did — logs of outputs, changes and deployments. Necessary, and not sufficient: the oversight articles turn on who. A platform log that says "user jsmith approved" proves an account clicked; it does not prove a competent, authorised person oversaw. Audit trails prove what the machine did. This infrastructure proves which authorised human oversaw it — to a standard a regulator can verify without trusting the operator. On the same clock, the platform is extending from what ships today toward regulation-as-configuration (the Act's requirements compiled as a versioned rails pack of hard policy floors, so regulatory change is a content update, not a rebuild), work-aware oversight posture (risk of the work classified live; assurance tightens automatically, tighten-only, reason always shown), and one-tap oversight transparency (the human-authority chain readable from the conversation itself) — each labelled honestly as in development, built during the sixteen months regulators just granted for exactly this.
Critical infrastructureAnnex III
Annex III lists AI used as a safety component in the management and operation of critical infrastructure — road traffic, and the supply of water, gas, heating and electricity — among the stand-alone high-risk classes now on the 2 December 2027 clock. For a utility, that class can reach systems that recommend or gate operational decisions: network-model "what-ifs", pressure and flow interventions, meter-driven anomaly responses. Where an operator concludes a system falls in scope, Articles 14 and 12 translate into operational questions — which person accepted the system's recommendation, under what authority and conditions, and can the record be verified independently, years later? A sealed human acceptance bound to the decision, with an audit chain a regulator can check without trusting the operator's own logs, is one mechanism designed to answer exactly that — and it retrofits: the operational system stays exactly as it is, and oversight adds at its decision points through simple APIs, no rebuild required. Classification is the operator's assessment to make — many utilities have compliance work well underway — and the mapping here is offered as a reference point, not a diagnosis.
Language discipline, stated plainly: this page describes how the infrastructure is designed to support compliance with the cited provisions. It is not a claim of certification, conformity assessment, or legal advice; obligations depend on system classification, role, and deployment context, which the deploying organisation must assess. We keep this wording strict because a trust product that overclaims its own compliance posture has failed at its one job.