Hardware we own and run
The big card.
brain is the largest card in strata→signal's own hardware — the one with room to hold the biggest models we run. RuleSage, amble and RealKeep send their model calls to these machines instead of a cloud service, and brain takes the jobs nothing else can fit. An upgrade the products can call on, never a dependency they're stuck with.
- depth on demand
- idle by default
- an upgrade, not a dependency
Today, one machine performs several roles. The fleet plan gives each its own box; the names move with the roles.
brain runs the models that check other models' work — it scored another model's answers on RuleSage's bake, the run that turns a rulebook into verified answers. The harder reasoning, the longer thread, the answer worth getting exactly right rather than merely close: that is what the big card is for.
Day to day the smaller models answer and brain sits idle. The idleness is the design, not waste — nothing here stops working if brain is busy or offline. It is the heavy option, never a load-bearing wall.
Today
The card is in, the chassis isn't.
brain isn't a separate machine yet. The card is installed in keel — the box our models run on — and brain's own chassis is bought but not assembled. Serving moved off this box on 13 August 2026: a rented server faces the public now, and reaches keel over a private tunnel for the generation. What stays here is the thinking. When the card moves into its own chassis, this page describes the same job, and nothing that calls it has to change. The name travels with the role, not the metal.
- as of 2026-08-14
Where this sits
keel is the box our models currently run in; brain is the card inside it. Two more hardware roles are named in the plan but their metal isn't assembled.
keel · RuleSage · amble · RealKeep · strata→signal