What the implementation does
For a Kimi-family model ID, the local registry records CN as the weights-origin country. A recognized endpoint selects a separate provider route whose defaults supply hosting jurisdiction and retention fields. The application can override some hosting and compliance fields. The SDK does not inspect model weights, locate physical hardware, or test a provider's retention behavior.
Two questions, one answer given
A configured hosting jurisdiction answers a different question from a configured weights-origin country. A retention setting is different again. Keeping these fields separate lets a receipt preserve exactly which values the SDK used for one decision without treating any registry value as provider-issued proof.
What AgentGuard records
The AgentGuard MCP server has a provenance_preview tool and a spend_decide gate. Both accept model metadata, not prompts. The preview is unsigned and records no spend. A successful spend decision can include the provenance block inside an Ed25519-signed, content-free receipt. These four classifications are reproducible from the current local registry:
The same model ID produces different route defaults. If signing is configured and spend_decide returns a signed decision, the signature protects those recorded fields from later editing. It does not independently prove the provider, physical hosting location, weights origin, or retention behavior.
Flagged is not banned
The flag is a policy input, not a judgment about a model. In the MCP server's default standard posture, spend_decide blocks registry-flagged foreign-origin weights unless the hosted consent endpoint accepts a supplied consent receipt ID. Compliance posture blocks them even with consent. After a valid standard-posture consent check, known-cost and configured cap checks still decide whether the call is allowed.
One honest caveat
The Fireworks, Together, Baseten, and Moonshot route entries in the supplied registry default to 30 retention days. An integrator can configure another value. A signed receipt preserves the value used by the SDK and its capture time; it does not verify the provider's actual retention behavior.
Try it
npx -y @agentguard-run/mcp
# then call provenance_preview with:
# model: kimi-k3
# endpoint_url: https://api.fireworks.ai/inference/v1
- model
- kimi-k3
- weights origin
- CN (registry classification)
- jurisdiction
- US (fireworks)
- retention
- 30 days (registry default)
Reproduce the classification with @agentguard-run/mcp or inspect the provider registry in @agentguard-run/spend. The values above are current registry output, not a provider certification. Sovereignty guide: agentguard.run/sovereign. Want a model or route added to the registry? [email protected].