A public option for AI.

A public option does not need market share to do its job. It needs to exist, because its existence changes what the commercial options can quietly do.

Terms change. Prices change. Models are deprecated on ninety days' notice, or replaced by a checkpoint with the same name and different behaviour. Services withdraw from jurisdictions, or are compelled to retain what passes through them. Each of those is survivable if there is somewhere else to stand, and none of them is survivable if the only capable models are ones you rent.

The floor is not "a worse free version". It is a model good enough to do real work, that cannot be withdrawn from you, because you have the weights and the means to make them again.

Encryption on the web used to be a paid, manual privilege. Certificates cost money and took a morning, so most sites went without, and the default for ordinary traffic was plaintext. Then certificates became free and automatic, and within a few years the default flipped.

The web did not get encrypted by asking companies to be more generous. It got encrypted because someone made the good option free, automatic and boring — and then the commercial market adjusted around the new floor.

That is the shape of the claim here. Not that open models will beat the frontier labs. That a capable, free, verifiable model changes what the default is.

So the whole of homespun is built to be reproduced rather than believed. Every input is pinned to a hash — the corpus at dataset commits with a licence recorded per document, the code commit, the environment, the configuration, the container digests. Every output is hashed. One signature covers all of it, and the verifier holds only a public key.

That matters more for a public option than for a commercial one. A company can be held to its claims by contract and by the prospect of losing your business. An institution offering something free has neither, so it should offer proof instead. What is in it, and how to check it.

The arithmetic is unglamorous and that is the point: six floating-point operations per parameter per token, at 40% utilisation on current hardware. Budget 30%.

A capable model at the size worth running costs less than a mile of urban arterial road, and rather less than a single year of a mid-sized agency's software licences. Infrastructure arguments usually founder on the number; this one does not.