> For the complete documentation index, see [llms.txt](https://osintelligence-llc.gitbook.io/osintelligence/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://osintelligence-llc.gitbook.io/osintelligence/the-sovereign-stack.md).

# The Sovereign Stack

**Architecture, Discipline, and Evidence from One Desk**

Jamey Kistner · OSINTelligence LLC

**TL;DR** — short on time? [**What is novel in this record**](/osintelligence/what-is-novel-in-this-record.md) is the one-page map: five headline claims, each with a sealed receipt, and every chapter's contribution linked. Start there.

***

## What this is

A scholarly monograph on sovereign AI systems: 24 chapters of architecture, method, failure record, and measured results, produced by a single operator, self-funded, on consumer hardware, with every claim receipted.

When NVIDIA's Jensen Huang defined the terms the industry now runs on, he called the modern AI datacenter an **AI factory**, a **token factory**, whose product is intelligence and whose economics reduce to one relation: output scales as tokens-per-watt multiplied by available power. That framing was built from abundance, optimizing inside grid limits the hyperscalers can no longer outbuild. **This record is the same architecture reached from the other end.** It is a functional AI factory arrived at from scarcity rather than capital: a full-stack co-design, one router hot-swapping a dozen models across a single 12 GB card, a mixture-of-experts model streaming its expert bulk through system memory while only its attention sits on the card, work dispatched as tokens across a governed fabric, the whole of it inside a hard power envelope on one desk. Two builders reached the same governing equation, tokens-per-watt times power, from opposite resource extremes; the fuller account of that convergence is in *Sovereign Sustainability* (Chapter 22).

The research runs on a discipline of pre-registered, sealed experiments: hypotheses are hashed and dated before a run, results are reported against them in full, and the failures are published beside the successes. The failure record sits in the middle of the book, not in an appendix, because it is the strongest evidence of integrity the record has.

This site is the browsable mirror of that monograph. Chapters appear here at their graduated versions; each page carries its source version in its footer.

## How to read it

Three routes through the body of work, laid out in the Reader's map:

* **The argument** starts at [Part I](/osintelligence/part-i-the-architecture.md), the architecture and the thesis.
* **The method** starts at [Part II](/osintelligence/part-ii-the-discipline.md), how one operator holds it together.
* **The evidence** starts at [Part III](/osintelligence/part-iii-the-evidence-what-broke.md), what broke, then [Part IV](/osintelligence/part-iv-the-evidence-what-worked.md), what worked.

## License and citation

The entire work is released [**CC BY 4.0**](https://creativecommons.org/licenses/by/4.0/): reuse, redistribution, derivatives, and commercial use are all permitted. The only ask is attribution to **Jamey Kistner, OSINTelligence LLC**.

In-series works are cited by title (for example, *The Sovereign Triad* §4.2).