Sovereign vault and service host

os-totebox is the archive layer of the PointSav family: one isolated vault per entity. It stores the records, runs the services that process them, and exposes nothing else. An entity is whatever needs a separate set of books — a person, a corporation, a real property, a project, a household. Each entity has its own os-totebox. Toteboxes do not share files, do not share users, and cannot see each other. They communicate only through the Diode, and only on command from os-console or Fleet aggregator. This article covers the services inside, the WORM discipline, the current host shape, a known persistence limitation, the compute tiers, and the freely transferable design.

What lives inside

Each os-totebox hosts a fixed set of services:

Every service above is a real, currently active crate, not a placeholder. service-fs is worth a specific caveat: its own companion integrity-anchoring service has been failing since 2026-08-01, and checkpoint signing is deliberately unset at this baseline — the ledger is running, but not yet fully hardened. Two functions sometimes assumed to be separate os-totebox services — a deep archive of immutable records and a financial ledger — are not backed by dedicated services today; the nearest existing artifacts are thin front-end views with no independent service identity behind them.

The WORM discipline

service-fs writes raw payloads directly to append-only block storage. There is no delete operation in the code path. ¹ A compromised service cannot overwrite history because the verb does not exist at the storage interface. This is the architectural enforcement layer for processing integrity and the WORM ledger discipline.

Every institutional record lives as an inert flat file — Markdown, YAML, or CSV — that requires no proprietary runtime to read decades later. A .yaml ledger or .csv register is universally readable by any text editor, on any hardware, in any decade. Data migration cost falls toward zero: the operator always holds the source in a form no proprietary software can lock. The WORM storage architecture and ledger architecture articles describe the technical implementation.

The host shape

"os-totebox" names three distinct things today, and keeping them apart matters for reading the rest of this section correctly.

The production path is a seL4-isolated Linux guest. A hand-written, dependency-free seL4 protection domain — real, bare-metal AArch64 code, not a simulation — has booted, performed IPC, and driven genuine VirtIO network I/O to reach the Doorman service over a real TCP handshake. Above that low-level milestone sits the actual hosting design: a Microkit-managed VMM (libvmm) boots an ordinary, unmodified Linux guest, and the os-totebox binary runs inside it as the guest's own init process. seL4's formally verified capability boundary provides the isolation guarantee at the hypervisor layer; the services inside the guest run unmodified. A real, purpose-built guest root filesystem — Ubuntu 24.04 arm64 ("noble"), assembled via debootstrap — backs this path; the glibc base is a deliberate choice, needed for FFI compatibility with the C++ knowledge-graph engine service-content depends on. This produces a real, complete boot image (loader.img, roughly 113 MB) that has been live-verified: a genuine boot, followed by an end-to-end inference round trip through the Doorman and back. The full target architecture is a seven-domain capability map — a supervisory watchdog-pd, and six service domains including service-fs-pd, network-pd, service-content-pd, service-people-pd, service-slm-pd, and service-extraction-pd — with one invariant enforced at build time: only service-fs-pd ever receives the block-device capability. Every other domain reaches durable storage only through it.

A NetBSD 10.1 image also exists as a transitional artifact, not the production target. It is real and built — a genuine NetBSD cross-toolchain pipeline (nbmakefs, nbinstallboot) produces a bootable image with a Veriexec binary-signing manifest — but it is a compatibility bridge on the path to the seL4 guest above, not a parallel destination.

A third, unrelated meaning is the plain single-process bundle that actually runs os-totebox locally today. Its own main.rs spawns service-content and the Doorman (slm-doorman-server) as two threads inside one ordinary Linux process — each service's router and business logic are untouched, this file only launches them together. This is what is deployed as the local-totebox.service systemd unit on the development workspace. It is a legitimate, production-intended packaging shape in its own right (the same single-binary-multiple-role pattern used by tools like Vault or Nomad), but it is a different thing from the seL4-hosted guest described above, and the shared name is a real source of confusion worth naming plainly.

A known limitation — data persistence inside the seL4 guest

The seL4-hosted guest's virtio-blk storage device is attached but is never actually mounted by the guest's own init sequence. In practice, this means a reboot of that guest currently wipes all of its data — a direct tension with this article's own "sovereign, persistent data vault" framing, and a limitation stated here plainly rather than left implicit. A related gap compounds it: the QMP-based graceful-shutdown path is documented as non-functional, because the guest declares no ACPI or power-button device for QEMU to signal — so a redeploy today hard-kills the guest with no checkpoint guarantee. Neither gap affects service-fs's own WORM ledger, which is a separate, host-level service (see above); it affects the data a seL4-hosted guest itself accumulates internally. Both are known, open items, not silent defects.

Compute tiers

os-totebox adjusts its behaviour to available hardware, following the model-tier discipline that governs SLM usage:

Freely transferable

Every os-totebox instance is intended to ship as a single, signed boot image. The operator picks it up and moves it between cloud providers, a private server, or bare-metal at their own facility. There is no host operating system underneath that holds the keys. This is the Sovereign Addendum's intended commitment in physical form: the running instance remains the operator's property in any environment.

See also

• OS family — eight operating systems, one substrate — where os-totebox fits in the eight-OS family
• console-os — the Command Ledger that connects to os-totebox and presents its state
• Fleet aggregator — the fleet aggregator that queries many Toteboxes at once
• Diode standard — the unidirectional protocol through which the Totebox communicates
• seL4 microkernel substrate — the kernel underpinning os-totebox's isolation guarantees
• Machine-based authorization — how pairing governs access to a Totebox
• WORM ledger design — the append-only storage discipline enforced by os-totebox