root@fr:~$ doctrine --sovereign-risk
Sovereign risk. The one thing a model's power never tells you.
Choosing an AI model on raw capability alone only looks at half the picture. Raw power says nothing about what actually matters: who can cut your access, under which law your data flows, and what you have left the day the tap is closed. Here is the doctrine.
Sovereign risk
Risk-adjust before you compare.
In finance you do not buy a raw return: you adjust it for risk. An 8% return that can collapse overnight is worth less than a stable 1.8% that no one can take from you. AI follows the same rule. A more powerful frontier model that a foreign state can cut without notice and without recourse is not a better choice: it is an unadjusted return. You gain stable, auditable access under your own control.
Raw return
the model's pure power, the one number everyone looks at.
Sovereign risk
the chance that access, rules or jurisdiction change without your consent.
Adjusted return
what remains once risk is factored into the calculation. This is the number a business decision rests on.
The operational grid
Five axes. One real score.
A model's real score is not a single grade: it is a grid. Each axis can swing a decision, whatever the headline power.
Quality
Real capability on your tasks, not a marketing leaderboard. What the model produces on your documents, in your language, on your cases.
Access reliability
Does access hold over time? A model unavailable one morning means a stalled agent chain and a broken client promise.
Governance
Guardrails, audit logs, human oversight. Can you explain, trace and justify every answer in front of a regulator?
Legal footprint
Under which law the model runs and your data flows. Access outside Europe exposes you to the CLOUD Act and to orders you do not control.
Data protection
Are your data shielded from foreign-state interference? The only solid answer is architectural, not contractual.
You start with this grid, then you compare. Not the other way around.
The three walls
Access to frontier AI is filtered by three walls.
The passport wall
filters by passport, new
The third wall lets no cheque through. Access is decided for you, based on your country, your sector or geopolitical criteria you can neither audit nor anticipate. It can be cut off overnight.
12 June 2026: on US government order, a frontier lab cut access to its new model (Fable 5) for all foreign nationals. Unable to verify everyone's nationality, it cut it off for the entire world.
Anthropic, 12 June 2026As long as someone else holds the switch, owning your AI in France is the only hedge.
A big cheque clears the first two. Nothing clears the third. The passport wall is the one no cheque can clear: when access is decided for you, on criteria you can neither audit nor anticipate, sovereignty stops being an option. It becomes the only hedge available.
Sovereignty decided agent by agent (BYOK)
Sovereignty is decided agent by agent, not account-wide.
Covering a whole account at the protection level of the most sensitive process is expensive and slows everything down. Covering at the weakest level exposes your secrets. The right granularity is the agent. Every agent defaults to our sovereign models: open-weight, hosted in France, with no US entity in the data path. For an agent with no sovereignty constraint, bring your own keys (BYOK) and point it at a frontier model. Sensitivity decides, agent by agent.
| Process | Model | Jurisdiction | Status |
|---|---|---|---|
| Drafting legal deeds | Sovereign model (open-weight) | France | + Sovereign |
| Contract analysis | Mistral | France (EU) | + Sovereign |
| Document search | Sovereign model (open-weight) | France | + Sovereign |
| Internal summary, non-sensitive | Frontier model (BYOK) | United States | ! Exposed, by choice |
The Sovereignty Map is not a PDF: it is a living governance view inside the Workspace. Per agent: which model, where, under which law.
Lucidity, not slogans
Sovereignty of execution, not of the model's origin.
The best open-weight models today are largely Chinese. Running them in France gives you sovereignty of execution and data - your data stays in France, under your law - but not sovereignty of the model's origin. It is a distinction that matters, and we own it.
Sovereignty of execution and data
The model runs on French hardware, with no US entity in the data path. Your data never leaves the territory, whatever the origin of the weights.
Mistral, the serious European engine
Behind the frontier on raw power, but fully European on jurisdiction and origin. The choice when legal footprint outweighs the last benchmark point.
Insurance bought knowingly
We name what each model covers and what it does not. The right trade-off is made task by task, knowing exactly what you are buying.
Someone else's switch
The firm that had built its entire production on access it did not hold.
A firm had automated most of its production on an agent chain wired to a single foreign frontier model. Everything worked: faster, cheaper, delighted clients. No one had asked which infrastructure that model runs on, and who can cut access. The answer was simple: not them. The day access changes - export policy, a unilateral decision, opaque criteria - it is not a feature that falls. It is production that stops. The lesson is not to flee frontier models: it is to know, for each process, which switch you hold and which ones someone else holds for you.
- An agent chain wired to a single model is a single point of failure.
- Access you can neither audit nor anticipate is not a given: it is a reprieve.
- Mapping your switches does not slow innovation. It makes it insurable.
Map your switches before someone flips them.
The Sovereignty Map runs each of your critical processes through the grid: which model runs it, under which jurisdiction, and what you risk if access closes. Every line that reads foreign frontier model is a switch someone else holds. Start by knowing which ones.