The platform governing its own AI (§V.6 / §VIII.5)

Model registry & governance

The differentiator: a governance-AI that obeys the governance rules it argues for. Every model is registered with a version and a published impact assessment before it may be used; the AI advises, a named human decides.

✓ Passes its own AI rules Every governing model obeys §VIII.5 / §XIX.3: open-source, reproducible, owned, and never a decider.

The algorithm register (§VIII.5)

System	Role	Owner	Impact tier	Audit	Decides?
§V.4 Router Classify who decides a question (value→people, technical→experts, interpretive→grounded). Routes, never rules.	route	Integrity & Safeguard Assembly	high	passed	never
§XIX.3 Threat-Matrix Detect candidate crisis signals fast. Authorises nothing; a human body declares.	detect	Crisis Defence Council (oversight: Integrity Assembly)	high	passed	never
§V.6 Decision-Support Strengthen evidence: modelling, scenario-testing, anomaly detection, plain-English translation. A named human owns the decision.	model	Domain expert board (per decision)	limited	passed	never

Published algorithmic impact assessment · sovereignty is a configuration choice (SOVEREIGN_BASE_URL), not a rewrite.

Registered inference models

Model	Version	Sovereign	Notes
claude:claude-opus-4-8	opus-4-8	frontier	Frontier reasoning model for the MVP; not sovereign — see the impact assessment for the §VIII.5 path.
claude:claude-sonnet-4-6	sonnet-4-6	frontier	Default model: lower cost, strong faithfulness; used for grounding and the eval lanes.
sovereign:open-weights	configurable	sovereign	OpenAI-compatible open-weights endpoint (vLLM/Ollama/self-hosted). The §VIII.5 sovereignty path — set SOVEREIGN_BASE_URL to run the core on an inspectable, jurisdiction-controlled model.
stub	1	sovereign	Deterministic test double — fully inspectable, no external inference.
cassette	1	sovereign	Recorded-response replay for the CI faithfulness gate.