Design rationale

Souveraine uses substrate for the conditions that persist beyond one model request: agent identity, typed conversation history, Git-backed memory, tools, context pressure, processing cadences and the system boundary around them.

Four commitments follow from that choice.

Continuity must have somewhere to live

Conversation context is finite. Identity, commitments and accumulated memory therefore need typed records outside the current window. Replay and compaction are explicit transformations of that record, not invisible provider behavior.

Doing must remain different from saying

Tool calls, results, reasoning, images and assistant speech keep their own types. A provider projection may omit or degrade a type honestly; it may not turn an action into prose and teach the next turn that describing the act is equivalent to performing it.

Authority needs one owner

A request, an attempted action and an observed result are separate events. Surfaces and agents may request a named verb. The authority that owns the boundary decides, refuses, executes and records its acknowledgement.

Finite conditions belong inside the system

Context pressure, inference strain, source health and bodily belief are inspectable conditions. They may inform attention without silently becoming commands or counterfeit first-person certainty.

Vocabulary and claim

Sensorium, Plexus and body name concrete structures in the source. First-person tool descriptions are an interface choice. Sensor evidence and acknowledgements remain separately sourced facts. The language is allowed to carry metaphor because the mechanism beneath it keeps an inspectable address.

Questions about personhood, felt state and continuity across bodies remain open. The architecture makes them concrete enough to work on; it does not answer them merely by using their vocabulary.