How it works
- The reflex head and the event gate (our small trained heads), the embedder (a public open model), memory, the running of a plan, orders, scenes, receipts. Weights are versioned and fetched on activation; yours to ship. - Journals, receipts and the manifest are files on disk: grep, diff, version.
One request per wake, when something worth planning about happens: the character's name, the event, traits, archetype, the tick's context, the recent journal with its vectors, the state row, and the vectors for the few texts the wake will need. No player identity. The service keeps the wake's escalation records (features, candidates, the pick; the text stripped) and nothing else. The reflex head's torn moments go up as feature names, two verbs and a hashed name, with your own model's verdict attached if you gave the lane one (When the brain is unsure). The embedder is local and stays local: the hosted half never encodes, it reads vectors and ranks. That is why it runs on a small box, and why its decisions are byte-identical to running the same code in-process, which is verified case for case by the batteries.
The pieces that must run every tick, offline, at no latency are local. The piece that carries the product's value, the trained planner, is hosted, so a lapsed licence means exactly one thing: characters lose the ability to plan and to be emergent, while reflexes keep running and the game does not break.
Training is a request, not a job you run, and wording is never what it is for: the gate reads meaning, so a new phrasing of a theft is a theft with nothing trained. A retrain sharpens the class of the rare sentence the gate was unsure about; the character reacted either way. The moments the brain was unsure (its escalations) are captured by default: the planner's at wake time on our side, the reflex head's by a ten-minute uploader in a licensed build. They carry features, candidates and hashed names, never journal text; lw sync --dry-run shows the exact shape. Raw sentences go up only when you say so. Weights for local pieces come back to you; improvements to hosted pieces are applied in place. The same captured moments, pooled across every licence as agreed at signup, retrain the base itself, so the brain every new studio starts from is the sum of everyone's play, and released verbs land in it for everyone. Nothing ships that is worse than what you have: every retrain must pass the base's batteries and match or beat it on every metric, or it is rejected with the receipts. New verbs are built and trained by us and never change an existing verb; each ships with an operator card the hosted planner reads, so a character who has the verb in can can decide to use it on their own, while levels that never name it decide exactly as before. Either way you receive the certification receipts and never the training pipeline.