Policy at the Speed of Inference
A Proposal to Governments on Building Compliance Infrastructure for the AI Era
Policy Proposal · AI Governance · June 2026
Every regulatory system ever built was designed for a world that moves at human speed. Laws are debated, drafted, consulted upon, amended, and enforced through institutions that operate on timescales of months and years. This is not a flaw in institutional design. It is an accurate reflection of the pace at which the world, until very recently, actually changed.
That premise no longer holds. AI systems are trained, deployed, updated, and scaled in cycles measured in weeks. A model that passes a compliance assessment on Monday may behave differently by Friday — not because anyone made a deliberate change, but because the deployment context shifted, the input distribution drifted, or a fine-tuning step altered a capability no one had mapped to a legal obligation.
The gap between regulatory speed and AI speed is not a gap that more effort closes. It is structural. And it is compounding — every month that passes without a governance architecture built for this reality, the distance between what the law requires and what the law can actually enforce grows wider.
This proposal is not about writing better laws. It is about building the infrastructure that makes existing laws enforceable. It is addressed to governments, regulatory bodies, and the institutions charged with making AI accountability real — not aspirational.
"Compliance cannot be a certificate. It must be a continuous condition — one that is monitored, verified, and enforced in real time, by systems purpose-built for the speed at which AI operates."
The compliance audit, the point-in-time certification, the annual review — these instruments were built for static systems with stable behaviour. Applied to AI, they produce something that looks like accountability but functions as a liability shield. A system can pass its audit in January and violate fundamental rights in March, with no mechanism to detect the delta.
Governments that rely on these instruments are not governing AI. They are issuing certificates to it.
Part I — The Asymmetry That Changes Everything
Regulatory frameworks rest on a tacit assumption: that those who must comply and those who must enforce operate at roughly comparable speeds. A factory can be inspected. A bank's books can be audited. A drug's effects can be tested before market entry. The subject of regulation is slow enough to be caught.
AI inverts this assumption. The deployer of an AI system can iterate faster than any oversight body can inspect. The model can be updated faster than any certification process can complete. The harm can propagate — at scale, across borders, to millions of affected people — in less time than it takes to draft a regulatory inquiry.
| Dimension | Current reality |
|---|---|
| Average AI regulatory rulemaking cycle | 12–18 months |
| AI model update cycles | Days to weeks |
| Scale of harm at deployment | Hours |
This is not an argument that AI cannot be regulated. It is an argument that AI cannot be regulated with instruments designed for a different pace of change. The asymmetry is the problem. Closing it requires that the tools of compliance operate at the speed of the systems being assessed — not the speed of the institutions doing the assessing.
There is a second asymmetry that receives less attention: the asymmetry of consequence between errors. In compliance work, a false positive — flagging something compliant as a violation — costs friction and remediation effort. A false negative — clearing something non-compliant as compliant — can cost fundamental rights, trigger regulatory penalties, and cause harms to the people the law was written to protect. These are not equivalent errors.
The accountability illusion
When a government issues a compliance certificate to an AI system it has not continuously monitored, it has not ensured compliance. It has transferred legal risk to a document. The document provides cover for the deployer. It provides nothing for the person harmed when the system drifts.
Part II — Three Structural Failures
Current AI governance frameworks fail in three structural ways. Each failure is distinct. Each compounds the others. None is addressable by incremental improvement to existing instruments.
Failure 1: The Accountability Dissolution Problem
Law requires a responsible subject — an entity that made a choice, bears consequences, and can be held to account. AI systems dissolve this requirement. The behaviour that constitutes a violation may emerge from a training process no single engineer designed, deployed in a context no developer anticipated, producing an outcome no one intended.
When a regulator finds a violation, the first question is: who is responsible? Without a mechanism to trace the violation through the technical behaviour back to the specific decision that caused it, the answer is: no one in particular. That answer is not compatible with accountability.
Governance requirement
Any AI compliance architecture must produce a traceable chain — from legal obligation, through observed system behaviour, to the specific decision point that created the violation, to the party responsible for that decision.
Failure 2: The Compliance Capture Problem
Accountability mechanisms corrupt. The moment you create a compliance industry, you simultaneously create a market for compliance theatre. The parties who most need rigorous assessment are the same parties who fund the assessment, control the framing of it, and have the strongest incentive to ensure it produces a favourable result.
In AI governance, this dynamic is amplified. A general-purpose AI tool used for compliance assessment can be prompted, manipulated, and steered — its response distribution is not fixed, and a sophisticated actor can shift it in the direction of a favourable verdict.
Governance requirement
Compliance assessment tools must be adversarially robust — structurally incapable of producing a verdict that contradicts their legal grounding, regardless of how the question is framed or how persistent the pressure. Character, not instruction.
Failure 3: The Jurisdictional Inequality Problem
AI governance, as currently practised, is a privilege of the most technically sophisticated regulatory jurisdictions. The EU has the infrastructure, the enforcement apparatus, and the economic leverage to compel compliance from global AI deployers. Most jurisdictions do not.
This creates a two-tier compliance reality: systems that would be flagged as non-compliant under careful technical assessment go undetected in jurisdictions that lack the tools to detect them. The result is that the people most protected from AI harm are those in jurisdictions with the most resources — not those with the most vulnerable populations.
Governance requirement
Compliance infrastructure must be accessible to every jurisdiction that has passed a law. The existence of a codified right must be sufficient to trigger enforceable assessment — regardless of the technical capacity or economic leverage of the jurisdiction enforcing it.
Part III — What Governments Must Build
These three failures are not individually addressable. They require a new category of governance infrastructure — one that did not exist before AI made it necessary, and one that cannot be assembled from existing compliance tools.
The Compliance Intelligence Layer
What governments must build — or mandate the building of — is a compliance intelligence layer: specialised AI systems trained specifically on the text of the laws being enforced, purpose-built to assess whether other AI systems are operating within those laws, and designed from their foundations to be adversarially robust, continuously operational, and jurisdictionally accessible.
What this is NOT
- —A general-purpose LLM given a legal system prompt
- —A compliance chatbot that discusses regulatory obligations
- —An audit tool that produces point-in-time certificates
- —A tool whose outputs can be shifted by persistent reframing
What this IS
- —A model trained specifically on the text of the laws it enforces
- —A system that constructs traceable chains from law to verdict
- —Continuously operational — not periodic, not point-in-time
- —Adversarially robust at the architectural level, not by instruction
The Auditability Requirement
A verdict that cannot be explained should not be rendered. In any jurisdiction where enforcement actions can be challenged, a compliance verdict that cannot be traced to a specific legal provision, applied to a specific observed behaviour, through a documented chain of reasoning, is not defensible in an adversarial proceeding.
The compliance intelligence layer must therefore produce not verdicts but arguments — four-part structures that specify: what the law requires, what the system did, why those two things are in conflict, and what a compliant path looks like.
The Temporal Validity Requirement
Laws change. Implementing regulations are issued. Enforcement guidance is updated. A compliance verdict based on legal text that has since been amended is not merely unhelpful — it is actively misleading. Any compliance infrastructure built for AI must treat temporal validity as a first-class concern: every output must disclose the legal corpus date it is grounded in, every indication of legal change must trigger a confidence reduction, and speculation about post-training legal content must be structurally prohibited.
Part IV — The Five Demands
These are not recommendations. They are demands — the minimum conditions under which AI governance can function as governance rather than as performance.
Mandate Continuous Assessment, Not Periodic Certification
Governments must replace point-in-time compliance certificates with continuous assessment requirements for AI systems operating at scale. A system certified compliant in January is not certified compliant in March. Any AI system that affects decisions about individuals — in employment, credit, healthcare, public services, criminal justice — must be subject to ongoing technical assessment, not periodic review.
Specific requirement
Regulation must specify continuous monitoring obligations, with defined re-assessment triggers including model updates, deployment context changes, and incident reports.
Require Traceable Accountability Chains
Every compliance verdict rendered about an AI system must include a traceable chain: the specific legal obligation engaged, the specific observed behaviour, the specific reasoning connecting them, and the specific party responsible for the decision that produced the behaviour. Without this chain, violations cannot be enforced.
Specific requirement
Deployers must maintain technical accountability documentation sufficient to reconstruct a full violation chain for any compliance finding. Regulators must have the authority to demand this documentation without prior notice.
Fund Jurisdiction-Neutral Compliance Infrastructure
Every jurisdiction that has passed a law governing AI has an equal right to enforce it. That right is currently theoretical for most of them. Governments must fund the development of open compliance intelligence infrastructure that any jurisdiction can use, regardless of their technical capacity or economic leverage.
Specific requirement
AI governance frameworks must include provisions for technical assistance and shared compliance infrastructure for jurisdictions without the capacity to build their own assessment tools.
Establish Adversarial Robustness Standards for Compliance Tools
A compliance tool that can be manipulated into a favourable verdict is worse than no compliance tool at all — it provides false assurance, which is a mechanism of harm. Governments must establish minimum standards for the adversarial robustness of any AI system used for compliance assessment.
Specific requirement
Compliance assessment tools must be certified against adversarial robustness standards before being used in any regulatory or enforcement proceeding. The standards must include red-team evaluation and character-level testing.
Treat Compliance Assessment as Public Infrastructure
The current model — in which compliance assessment is a private market service sold to the parties being assessed — has a structural conflict of interest that no amount of ethical guidance resolves. Governments must treat AI compliance assessment as public infrastructure: independently funded, operationally independent from the parties being assessed, and accountable to the public rather than to its clients.
Specific requirement
National AI governance frameworks must establish independent compliance assessment authorities with dedicated technical capacity, protected from industry capture through funding independence, appointment processes, and statutory mandate.
Part V — The Decision Before Governments
Every government that has passed an AI governance law now faces a decision that the text of the law does not resolve: whether to enforce it as written, or to accept a version of compliance that provides the appearance of accountability without the substance of it.
That decision does not require a press announcement. It is made, quietly, in choices about what assessment infrastructure to fund, what standards to require of compliance tools, what independence to give assessment authorities, and what weight to assign to the distinction between a system that passed a certification and a system that is actually compliant.
| Dimension | Current practice | What governance requires |
|---|---|---|
| Compliance model | Point-in-time certification | Continuous technical assessment |
| Accountability chain | Narrative audit report | Traceable law → behaviour → verdict → remedy |
| Assessment tools | General-purpose LLMs with legal prompts | Specialised compliance intelligence models |
| Jurisdictional access | Dependent on technical capacity and resources | Public infrastructure accessible to all jurisdictions |
| Institutional independence | Market-funded assessors chosen by deployers | Independently funded public assessment authorities |
| Error asymmetry | False positives and negatives treated equally | False negatives treated as categorically worse |
The Stakes of Getting This Wrong
The stakes of getting this wrong are not abstract. They are the rights of specific people — people who will be denied employment by an unmonitored algorithmic system, people whose data will be processed without their knowledge or consent, people who will be subjected to automated decision-making that no one is accountable for, in jurisdictions where no one has the technical capacity to detect the violation.
The exponential acceleration of AI capability means that every year of governance delay compounds. The systems that will be deployed in three years' time will be more capable, more widely used, and more consequential than those deployed today. The governance architecture that is built — or not built — in the next two years will determine whether those systems operate within the law or merely carry documentation claiming they do.
The minimum viable position
A government that has passed an AI governance law but has not funded the technical infrastructure to assess compliance with it has not governed AI. It has governed the appearance of AI governance. The minimum viable position is not a law — it is a law plus the machinery to enforce it.
The policymaking process was built for a slower world. That is not a criticism — it was an accurate design choice for the world that existed. The world has changed. The design must change with it.
Policy at the Speed of Inference means building compliance infrastructure that operates at the speed of the systems it governs — continuously, traceably, robustly, and in every jurisdiction where the law says rights exist. Not someday. Now. Before the gap between what is legally required and what is technically assessed becomes a distance that no enforcement action can close.
The law does not slow down because the technology is fast. The technology does not slow down because the law is slow. Something has to close the gap. That something is infrastructure — and it has to be built deliberately, by governments, before the window closes.
Veesta Engineering and Compliance Team · June 2026