CODLAB / Blog · Software & development

Desktop Data Sovereignty: Atomic OS Images, Sandboxed Portals, and Opt-In Local AI in CardinalOS

14 Nov 2026 CODLAB Team 8 min

November 2026 · CODLAB Editorial Team

Short answer: Genuine data sovereignty on a personal computer comes from verifiable systems architecture, not marketing slogans: for the in-development CardinalOS desktop, we are designing a Fedora Atomic OS image (bootc) with reboot rollback, sandboxed Flatpak XDG portals, local backup workflows, a zero-telemetry design goal, and a disabled-by-default local help architecture.

1. When Your Own Desk Starts Watching Over Your Shoulder

For decades, personal computers honored a clear division of labor: the operating system booted hardware, isolated processes, and safeguarded local files. Across several commercial desktop platforms, that balance has shifted — for example through prompts pushing online accounts during first setup, promotional surfaces in system menus, background telemetry, or cloud assistants woven into file search and text entry.

For independent professionals, engineering teams, legal and medical practices, researchers, and privacy-conscious households, this creates a practical goal: using modern hardware and dependable everyday apps without negotiating local file actions with an external cloud platform. With CardinalOS, we are designing a desktop currently under construction (no publicly installable OS image exists yet) built around four target architectural layers.

2. Planned Foundation: Fedora Atomic and bootc Instead of Fragile Package Drift

A common question among people evaluating Linux for daily work is update reliability: “What if a system update breaks graphics drivers right before a deadline?” The CardinalOS target architecture addresses this structurally through Fedora Atomic Desktops and transactional container OS images (bootc; this OS base, its signing pipeline, and public update keys remain under construction and are not yet shipped as a downloadable ISO).

  • Planned background staging: In a bootc architecture, updates do not overwrite libraries inside a running session mid-work; instead, a complete system image (to be cryptographically signed in the release pipeline) is staged alongside the active deployment and activated on reboot.
  • Deployment switch on reboot (rollback boundary): If a new kernel or graphics driver causes a regression on a machine, bootc lets the user boot the prior system deployment under /usr. Switching deployments does not rewind mutable state in /var, personal files, or external services (we do not publish reboot-time figures before measuring a release image on target hardware).
  • Separation between base OS and mutable workspace: Following the bootc filesystem model, the base OS under /usr remains read-only while system configuration in /etc and user home/application data under /var stay separate (disk encryption options belong to the future installer workflow and are not claimed as an already-delivered system state).

3. Sandboxing Boundaries, Portals, and Backup Design

On a traditional desktop, an unconfined utility generally inherits the read permissions of the logged-in user account. In the CardinalOS design, everyday applications target Flatpak sandboxing paired with XDG Desktop Portals — with explicit technical boundaries:

  1. Portal file access for sandboxed apps: When a sandboxed Flatpak app without broader static filesystem overrides opens a document through the FileChooser portal, it receives access to that selected file rather than the entire home tree. Unisolated apps or broad static filesystem permissions are not automatically revoked by portals, and full permission controls in the Settings panel remain in integration.
  2. Portal dialogs for screen and camera, plus a separate microphone path: For sandboxed apps without static device or socket overrides, the XDG ScreenCast portal (screen and window sharing under Wayland) and Camera portal (camera video nodes via PipeWire) mediate access through explicit selection and confirmation dialogs. By contrast, upstream XDG Desktop Portal does not expose a general microphone portal because Flatpak audio capture also runs through audio socket permissions; attributing prompt-based or fine-grained microphone controls therefore requires dedicated implementation and verification prior to release.
  3. Local Btrfs snapshots and planned restic backups: In the target storage design, local unencrypted Btrfs filesystem snapshots for local restore points are paired with a separate deduplicated, client-side encrypted backup workflow via restic to an external drive, local NAS, or user-chosen European storage target.

We apply the same precision to telemetry: excluding background OS telemetry, advertising IDs, and mandatory first-boot online accounts is a core CardinalOS design goal. Because the OS image, full package inventory, and network-filtering rules are still being completed, we state this as an architectural target rather than a finished traffic measurement; any external mail, sync, store, or web services configured by the user follow their own providers’ policies.

4. Local Help Instead of an Uninvited AI Copilot: Inside Cardinal Intelligenza

As an engineering studio focused on sovereign AI and grounded retrieval, it might be tempting to place an AI writing button inside every toolbar of CardinalOS. We deliberately took the opposite route: unconstrained text rewriting and uninvited assistants are excluded by design.

Many people switch operating systems precisely because they do not want an assistant rewriting emails, capturing screens, or feeding private documents into a language model. Assistance is helpful where a user wants to understand a specific system setting, inspect an error code, or locate a file on their own disk — without personal contents flowing to a model or over the network.

That boundary defines the candidate protocol for Cardinal Intelligenza (local Help, internal engineering documentation, not yet public) around four rules:

  • Disabled by default with deterministic fallback: Without explicit activation, zero model processes occupy RAM or GPU; local Help still works deterministically against verified local documentation when the model is off, and no conversation history is saved.
  • Extractive system Help rather than reading private files: Screenshots, histories, excerpts, and user documents are refused before the model starts. The model only selects a reference ID from local Help sources (Settings and system contract); the service extracts the exact text from SHA-256-verified local source bytes (mirroring the separation in Evaluating RAG: retrieval and answers separately) or replies honestly “I don’t know”.
  • Strict separation between local file index and language model: Local file lookup runs as an independent index outside the language model: restricted to explicitly chosen folders under single-use consent per indexing run, covering filenames by default and plain text only via a separate opt-in (no photos, no messages, and no private PDFs sent to the model). Both searching and opening a result require separate user confirmation, and filenames, paths, and indexed words never reach the model.
  • Honest development status (November 2026): The local llama.cpp / GGUF Help protocol is verified in an isolated test harness using a synthetic engine double and is not yet deployed system-wide. Neither model weights (downloaded only after explicit consent) nor real CPU/GPU latency and memory measurements on consumer hardware have been completed yet.

5. Sovereignty Means Future Verifiability

Data sovereignty is not a badge printed on a landing page. Our standard for a public release is that users, administrators, and independent auditors must be able to inspect the shipped OS image, its upstream open-source components, and the network endpoints a machine contacts.

While CardinalOS is in development, we separate what is already verified in laboratory harnesses from what remains to be built before launch — across the “Nitida” token pipeline, lab-tested continuity protocols, and the planned atomic OS foundation. You can review captures from isolated development sessions on the CardinalOS product page and request a single availability notice at cardinalos.org.

Frequently Asked Questions (FAQ)

What is the CardinalOS design goal regarding telemetry and online accounts?

CardinalOS is designed around zero background OS telemetry, zero advertising identifiers, and no mandatory online account during setup. Because the OS image is not publicly delivered yet and network-filtering coverage remains in progress, we state this honestly as a target design contract rather than a completed traffic audit.

Will CardinalOS run an AI assistant sending my files to a language model?

No. «Cardinal Intelligenza» is engineered as opt-in local system Help: disabled by default, shipping without preloaded model weights, and refusing personal documents or screenshots before model startup. The separate local file index covers only explicitly chosen folders under single-use consent (filenames, plus optional plain text) without passing personal data to the model.

How is the planned update model designed to handle driver regressions?

The planned Fedora Atomic (bootc) foundation is designed to stage updates as a complete new /usr image alongside the active deployment for the next reboot (not yet published as an ISO or public signing chain). If a regression occurs, rebooting into the prior /usr deployment switches the system layer without rewinding mutable user files in /var.

Share this article

Comments

No comments yet. Be the first to comment!

Leave a comment

Will not be published

Related Articles

Project on your mind?

We develop tailored software and AI solutions. Let's start your next project together.

Free initial consultation