The evidence

We did not invent the claim that this is possible. We measured it against the field, and the field agrees.

It is viable. OpenJarvis (Stanford, 2026) showed that a decomposed local stack — intelligence, engine, agents, tools, memory — can reach within 3.2 points of a frontier cloud model on personal-AI benchmarks, at roughly 800× lower marginal cost, with inference on the device. Minions (ICML 2025) showed the raw context can stay local while a frontier model only orchestrates, at ~98% of the quality for 5–30× less. The gap between "local" and "cloud" is a systems problem, and it closes.

It is necessary. PrivacyBench (2025) measured retrieval-augmented assistants leaking a user's secrets in 16% of conversations by default. Opal (2026) showed that even the pattern of retrieval leaks. Scoping is the distance between a home and a sieve.

It is open ground. Multi-tenant community machines — friends and family sharing one computer and one AI — are almost unstudied; the literature is all enterprise Kubernetes. Cloudflare's own Mesh announcement (April 2026) admits that per-agent identity at the network layer remains unsolved industry-wide. The UK's ARIA is funding millions to verify the control plane of a WireGuard overlay. We are standing in a space the field has named and left empty.

[FIGURE 5 — "The white space": a map of what is studied (enterprise tenancy, cloud privacy) and what is empty (community tenancy on an overlay). Sovereign's mark in the empty space.]