Decisions this project actually made

Posts about decisions this project actually made and about what it measured. Every claim links the source it comes from; where a source states nothing, the post says so instead of filling the value in.

LLM EU Verified tarkoittaa dokumentoitua, ei oikeudellisesti sertifioitua. Se ei ole AI-asetuksen mukaisuuden arviointi, eikä LLM EU ole EU:n elin.

  1. Twelve green modules, and one assembled product

    Twelve owners, twelve passing suites, and a coordinator that enforced almost nothing. What the assembled product caught, and the rules that came out of it.

  2. What "the backend is a mock" means here

    The only runnable backend is a labelled mock. That label is a design constraint: what it has to say, what it must never claim, and which limits stay true because of it.

  3. Credits are non-transferable, and paid only for a verified pair

    Why the ledger has no transfer method, why a single unreplicated result pays nothing, and how the amount a device earns is actually computed.

  4. A licence attaches to a variant, not to a family name

    Teuken-7B names two differently licensed models on one 7B base. Two model cards, two verdicts, and the registry defect that only a second repository caught.

  5. Residency is a filter, never a preference

    A routing rule that can be relaxed under load is not a residency guarantee. This is the failure mode the policy refuses to allow, and the measured change that found it.

  6. Why the swarm sends a whole request to one device, not a model across many

    The popular image of a compute swarm is one model split across many machines. The arithmetic of consumer links is why this design does the opposite, and what it buys instead.

The compute swarm blueprint: what actually ran, and what did not