Infrastructure, not a product.

A checkpoint on its own is not infrastructure; it is an artefact that happens to be downloadable. What makes this maintainable by someone other than its authors is that the inputs are pinned and the outputs are hashed, so the thing can be rebuilt, audited, forked and handed on.

A company that publishes a model has a reason to hold something back: the corpus is the moat, the evaluation is the marketing, and the weights are the product. None of those incentives are dishonest, and all of them argue against publishing the one thing that would let you check the claims.

An organisation whose mission is the artefact itself has the opposite incentives. There is no reason to withhold the corpus, no advantage in an unreproducible evaluation, and a positive obligation to make the thing outlive its funding. That is the shape this work should sit in, and it is why the pipeline is built to be run by people who are not us.

It also changes the failure mode. A product that stops being commercially interesting disappears. Infrastructure that stops being commercially interesting gets maintained anyway, by whoever still needs it — provided they were given everything required to do so.

The build needs a conformant Kubernetes cluster and nothing else: no operators, no custom resources, no private registry, no vendor account. Four public upstream images pinned by digest, and every line of project-specific code shipped as configuration data.

That is deliberate, and it is what makes sovereignty a practical claim rather than a rhetorical one. A health system, a ministry, a university or a library consortium can run the same build on its own hardware, in its own jurisdiction, and verify that what it produced matches what was published — without asking permission and without a support contract.