Blog ai-security

The world's first government demonstration of AI compute verification

Amodo and the Dutch government ran the first AI verification experiment on government compute.

Amodo Design Team

6 November 2026

Contents

The Amodo recomputation processor in an open flight case on a data-centre floor, cabled to a laptop showing the verification dashboard — Figure 1. The recomputation processor in the Dutch government test lab. The verifier’s results are read out on the laptop.

In September 2026, Dutch Prime Minister Rob Jetten presented a research agenda for AI Compute Verification at a UN General Assembly event. Following this first public international discussion on AI verification, the Dutch government have been the first government to run a public test of an AI Verification system on government-run compute.

An Amodo engineer walks State Secretary Willemijn Aerdts and officials through the verification system on a laptop — Figure 2. State Secretary Willemijn Aerdts (Digital Economy and Sovereignty) is walked through the verification system.

Technical tools that enable coordination over AI development are often called “AI verification technologies”. AI verification technologies enable us to make claims about AI development and deployment provable, rather than trusting assertions and documentation.

About AI Verification — excerpt from “Research Priorities for AI Compute Verification”

Advanced AI poses large-scale risks that threaten the national security of every state. These risks cross borders, so no state can manage them through domestic measures alone. States have historically embedded the rules for cooperation on such risks in international agreements.

Where the stakes are high, the scope of those agreements is determined by what can be verified. For example, negotiators sought a comprehensive nuclear test ban but could not agree on how to detect underground tests. The Limited Test Ban Treaty excluded those, prohibiting only what could be verified through national technical means at the time: atmospheric, underwater and outer-space tests. By contrast, strengthening the Biological Weapons Convention has been handicapped by verification challenges. The presence or absence of a credible verification mechanism has been among the strongest determinants of whether a regime achieved its objective.

Compute is a promising option for AI verification because it’s indispensable for AI development, very difficult to manufacture, and could potentially make automatic claims about the computations which developers run through them. This could potentially be done by, for example, using privacy-preserving proofs that demonstrate compliance without revealing confidential information or disrupting operations.

What was the experiment?

When we talk about AI verification, we refer to a prover running a system and a verifier who is checking that the system meets an agreed specification or set of rules. In this setup, the Dutch government worked as the prover and Amodo as the verifier.

The rule being checked was that “this cluster may only serve Qwen 2.5 7B”. This is a specific operationalisation of a few different objectives that companies and governments may have:

Specific data centres are only being used for inference and not training. → This system does that by checking that all packets are consistent with an inference, not training, workload.
Only models approved (e.g. following the White House’s framework) can be served. → This system can check that tokens come from a specific approved model.

On 27th October, an Amodo engineer arrived at a Dutch government test lab in The Hague with an inference verification (recomputation) processor. Under the supervision of the SSC-ICT, Amodo installed the recomputation hardware. This processor receives network packets from the NVIDIA GPU server being run by the Dutch government. In this early experiment, the packets were forwarded using a prover-controlled network switch. In future instantiations, this system will use a mutually trusted network certifier, reducing the risk that the prover is able to cheat, whilst keeping its information confidential.

Close-up of the circuit board inside the recomputation processor — Figure 3. Inside the recomputation processor.

As part of the experiment, the prover ran both the approved model (Qwen 2.5 7B), and an “illegal” model, Qwen 2.5 1.5B.
The demo allowed users to send prompts to both models and see where the verification system flagged “allowed” and “illegal” model responses.

The verification dashboard flagging a response as FAIL, with a 58% token match against the verifier’s recomputation — Figure 4. A response claimed as Qwen 2.5 7B but served by the “illegal” Qwen 2.5 1.5B. The verifier’s recomputation only matches 58% of its tokens, so it is flagged as a fail; responses from the approved model match 100%.

What is next for this work?

This experiment showed the principles of a simple AI verification system, but there is a lot more work to do to make this both mutually trustable and adversarially robust. The two important next steps are:

Running verification using Zero-Knowledge Proofs. To preserve the confidentiality of the prover’s IP, verification systems should aim to use zero-knowledge proofs, cryptographic systems that prove the inference without revealing any IP. Amodo already has an early demo of this, and other research groups have advanced systems (e.g. Attestable).
Building a mutually trusted network certifier. Should AI verification require hardware to be installed in the prover’s cluster, it will need to be both trusted by the prover and hardened against tampering by the prover. Both Amodo and FLI are working on formally verified and open-source network certifiers.

In addition to work to extend this verification system, there are lots of other promising research directions: software-only verification (ZKPs), verification using analogue sensors, memory and process accounting. Amodo is not alone in working on these technologies; startups and non-profits in the US, Europe, and Singapore (Lucid Computing, Attestable, SASH, GPAI, FLI) are all working on AI verification R&D.

Future experiments

As AI verification systems develop, so too does their testing. Startups, non-profits, and governments developing these tools need access to larger clusters and nation-state-level red-teaming to test the suitability of the technologies developed.

In the summer of 1988, the US and Soviet Union ran the “Joint Verification Experiment”. Adversaries worked together to test whether their verification tools could detect one another’s nuclear tests. To build trust, AI verification tools will need to undergo similar Joint Verification Experiments, both between governments and between companies.

Amodo, Dutch government officials and an NRC journalist around a table discussing the experiment — Figure 5. Discussing the experiment with staff from SSC-ICT and the Dutch Ministry of Foreign Affairs, including Wouter Jurgens, Director of Security Policy, and with Hugo Schiffers of the Dutch newspaper NRC.

Quotes

State Secretary Willemijn Aerdts (Digital Economy and Sovereignty, D66) calls it a “world first” in the field of AI. Not in the frontier labs of American tech companies or their Chinese counterparts, but in a government office on the Koningskade in The Hague. It is repeatedly emphasized that a special experiment is taking place here, on this Wednesday afternoon in late October.

A system capable of monitoring artificial intelligence is shown. This makes it possible to “supervise what is happening in large data centers,” according to Aerdts.

— Excerpt from Hugo Schiffers, NRC, 5 November 2026 (translated from Dutch)

Why is today’s demonstration a first? “It is the first time that this type of ‘hardware verification’ is being tested on a government server,” says Wouter Jurgens, Director of International Security Policy at the Ministry of Foreign Affairs, who also came to observe the experiment.

Previously, the technology was only tested on Amodo’s own systems. “You can understand that governments are generally quite reluctant to act as guinea pigs for new equipment,” says Jurgens. “But that shows how urgent the situation is.”

— Excerpt from Hugo Schiffers, NRC, 5 November 2026 (translated from Dutch)

“This first AI verification experiment run on government compute marks the start of an important and ambitious project to verify AI development and deployment. We cannot develop these technologies alone; government partnerships, like this one with the Dutch government, are essential to making sure AI verification succeeds.”

— Thomas Milton, CEO, Amodo Design