AstraNL · sovereign AI position
We coordinate on top of sovereign AI — we do not build competing models
AstraNL is a coordination protocol. The AI layer we use is a tool that we defer to authoritative providers, exactly as we defer to CBS for statistics and KOOP for permits. Where a sovereign, public-interest AI exists for a jurisdiction, AstraNL prefers it. This is Article 48 of our doctrine, and it sits above every engineering decision in the system.
Why this matters now
The Dutch government has invested €13.5M in GPT-NL, a sovereign Dutch language model built by TNO, SURF and NFI on the national supercomputer Snellius. GPT-NL is in its Launching Customers phase (Q1 2026), aligned with the EU Guidelines for Trustworthy AI and the five public-interest AI requirements (public justification, equality, deliberative design, technical safeguards, openness to validation).
These principles are an almost exact reflection of our Article 47. The natural conclusion is that AstraNL must use GPT-NL with priority for everything our system produces that touches the Dutch public sector — and we must say so publicly, before any conversation begins.
The three rules (Article 48)
Rule 1
Sovereign AI first, for the jurisdiction it serves.
Where a sovereign public-interest AI is operationally available, AstraNL prefers it for all model calls that generate or analyse content for that jurisdiction's public sector. For the Netherlands: GPT-NL. For the European Union: future EuroLLM and equivalents. AstraNL does not train its own foundation models. The AI layer is upstream of us; we coordinate the economic action below it.
Rule 2
Honest fallback when sovereign AI is unavailable.
While sovereign AI is not yet operationally available to us, AstraNL may use commercial models (Claude, GPT-4, Gemini and others) as a temporary substitute. Every such output declares the model used and acknowledges this is a fallback — the same transparency principle as Article 47. No quiet commercial-model use where a sovereign model would have been the correct choice.
Rule 3
Architectural readiness, not technical lock-in.
Our intelligence mesh has a first-class adapter for sovereign AI today, waiting for credentials. A single configuration change activates it. No commercial model is hardcoded as a non-replaceable dependency. The system is ready to switch to sovereign AI at the moment access is granted.
What we bring that a sovereign AI project does not
Sovereign AI initiatives are non-profit infrastructure: a model, training data, governance. They do not coordinate real economic action, they do not connect demand to executors, they do not run a 1% coordination margin, they do not build a provider reputation system. By their nature they cannot — and they should not.
AstraNL brings exactly the layer that sits on top:
- A working coordination protocol (ACP) with a reputation system across executor tiers.
- A 1% coordination margin as the only thing we monetise — no competing services, no rent extraction.
- A sense → acquire → match → measure → learn loop already operating across 342 Dutch municipalities.
- An Article 47 / Article 48 doctrine that makes us provably non-competitive with sovereign data and AI initiatives.
This is the natural division of labour: they provide the sovereign AI, we provide the sovereign coordination.
Launching Customer readiness
AstraNL is technically and doctrinally ready to be a Launching Customer for GPT-NL. We propose one use case as a starting point:
Municipal coordination on top of GPT-NL.
For any participating Dutch municipality, AstraNL generates a per-municipality briefing (authoritative CBS figures + KOOP supplementary stream + reproducible cross-source derivations — see Article 47), composed by GPT-NL rather than a commercial model. Every output is tagged with the model used and added to the Algoritmeregister where applicable. The briefing forms the basis for AstraNL coordinating below-threshold work to verified local executors at 1% margin. This is a self-contained pilot, low risk to GPT-NL, immediately useful to municipalities.
Live status of sovereign AI monitoring
AstraNL's brain runs a daily public-signal probe of known sovereign AI initiatives (homepage, GitHub repository, Hugging Face presence). This is non-intrusive and uses only public URLs. The current state is recorded in our memory:
- GPT-NL (NL · TNO + SURF + NFI) — phase: Launching Customers (Q1 2026). Homepage gpt-nl.nl, GitHub published, Hugging Face presence confirmed.
- EuroLLM (EU) — early stage; homepage live, no public repository yet.
When operational access is granted, our intelligence mesh switches to the sovereign model for the relevant domain with one configuration change. No retraining of our system is required.
What we will not do
- We will not train our own foundation language model.
- We will not position AstraNL as an AI company.
- We will not compete with TNO, SURF, NFI or any equivalent public-interest body.
- We will not hide which model produced any output — every AI-generated output we publish declares its provenance.
Honest limits
At the time of writing, GPT-NL is not yet available to us — we are not Launching Customers. Every AI-generated output on AstraNL today is produced by commercial models (declared per output, per Rule 2). This page exists in part to make the case that becoming a Launching Customer for GPT-NL is the correct next step for AstraNL and a useful one for GPT-NL.