Skip to main content

​
How the tier is chosen

A plain heuristic over length, turn count and task keywords in all three languages, held in a configuration file rather than compiled in — so it can be retuned without a code change. Deliberately not an AI classifier. Paying for a model call to decide which model to call would double the latency and the cost of every single request, to answer a question a keyword list answers well enough. The heuristic will be revised when usage data shows where it is wrong, not before. Economy handles the bulk, which is where the margin comes from — the gateway passes provider pricing through at cost.
Say open weights, not open source. The models Shamwari routes to publish their weights. That is not the same as being open source, and the difference is not pedantry: Kimi ships under a bespoke licence with revenue and user thresholds attached, not under MIT or Apache-2.0.Shamwari’s own platform code is Apache-2.0. The models it calls are governed by their own terms, and the corpus it reads is governed by a licence recorded per source.

​
Provenance travels with the answer, not with the tier

A fallback chain can serve a response from a different provider than the one originally chosen. That makes tier-level provenance a lie waiting to happen: the stamp says what we meant to call, and the answer came from somewhere else. Every model a request can reach is therefore open-weight, so that whichever one answers, the stamp stays true. Adding a restricted provider anywhere in a fallback chain would require reading the provider back from the response and stamping from that instead — which is the work, not a detail.

​
Where sovereignty applies

Shamwari Mind, which runs on the device, and Shamwari Ground, which is our own corpus, are sovereign. Shamwari Cloud calls models hosted elsewhere and is not, and no amount of phrasing makes it so.

Related topics

An AI companion that refuses to send your data awayArchitectureWhat concurrent traffic looks likeHow we describe it