US and Chinese models, sovereign at runtime

The claim, and its boundary

A US- or Chinese-origin open-weight model, running on a verified EU/EFTA endpoint, is sovereign at runtime. It is not sovereign in origin, and NeuralRing never says it is. These are two different axes, and NeuralRing states both, always, separately.

Runtime sovereignty is about where inference executes and under whose jurisdiction — the operator, the data centre, the key custody, the logs. CLOUD-Act-style compelled access attaches to the operator, not to where the weights were trained. Origin is about who developed the model and under whose jurisdiction, its license, its export-control exposure, and where its weights were trained and under what terms they are released. Origin never moves an endpoint's tier or assurance, and tier never launders origin.

A US-origin open-weight model on a verified German endpoint is genuinely sovereign at runtime. It is not the same thing as an EU-origin model on EU infrastructure, and NeuralRing's catalog, API, and attestation never let the first be read as the second.

What runtime sovereignty does NOT cover

Origin: a US or Chinese developer trained the model — recorded and shown, never hidden or neutralised by where it runs. Training data: what the weights were trained on is outside what NeuralRing observes or attests. Future-version supply: runtime sovereignty protects the weights you already hold in EU custody. It does not guarantee that the developer releases the next version, or releases it under terms that permit hosting — a future open-weight release can still be restricted, licensed differently, or withheld before it is published. Continuity for what exists, not a supply guarantee for what doesn't yet.

The worked example: the June 2026 Fable-5 export directive

On 13 June 2026, a US export directive barred non-US persons from access to a US closed model (Fable 5) — the directive also reached its underlying model, Mythos 5, so it covered the whole family. For seventeen days that model, reachable only through its vendor's US-hosted API, could be — and for affected users was — switched off at a layer no EU gateway sits in front of. The directive was lifted on 30 June 2026, but the lesson does not depend on the reversal.

A closed model behind a US API can be shut off by directive: no EU gateway, residency promise, or contract insulates a non-US person from an instruction issued to the US vendor at the API layer. This is why NeuralRing does not federate closed US or Chinese model APIs as supply; such models may be listed for comparison (Tier 3, never routable) — transparency, not supply. Already-released open weights, in EU custody, cannot be shut off the same way: a directive can restrict a vendor's service; it cannot reach into weights already lawfully resident in EU object storage and serving on an EU operator's hardware. But future versions can still be restricted before release — runtime sovereignty kept the June-2026 scenario's already-imported supply serving; it would not have produced a successor the developer chose to withhold.

How NeuralRing states the two axes

Everywhere a model appears, two separate, non-interchangeable signals are shown: a runtime signal (tier + assurance) and a colour-distinct origin signal (origin, weights class, license). They never share a colour system or facet group; nothing is painted green unless it is certified. An unverified origin claim is amber, a verified one is blue — exactly as on the runtime axis; green is reserved for certified. A closed_api entry renders as a transparency listing — Tier 3, never routable, never green.

What we do and don't claim versus other gateways

NeuralRing does not claim a feature another gateway lacks. Residency, in-region routing, and zero-retention are real and increasingly common. The claim is narrower and structural: NeuralRing proves, per request, where inference ran and which model answered, and it holds the weights in EU custody so that already-released open-weight supply keeps serving even when a vendor's service is directed off. Where another gateway asks for trust, NeuralRing hands you a record you can check yourself.