Plain-language glossary
Vocabulary, not marketingThe words that determine whether something can act.
SkyTwin’s terms carry operational meaning. This glossary points to the deeper source and public guides where a word affects behavior or a release claim.
Decision terms
| Term | Meaning |
|---|---|
| Signal | An input the system can interpret, such as a connector event or local source record. Its origin matters. |
| Decision context | The typed situation enriched with a twin profile, policies, history, and relevant constraints. |
| Candidate action | A structured possible response. It must include risk assessment and reversibility; it is not permission to execute. |
| Decision outcome | The selected result: admitted execution, a request for approval, or an explained non-action. |
| Explanation record | The human-readable account of what happened, which evidence and preferences mattered, and how to correct it. |
Safety terms
| Term | Meaning |
|---|---|
| Trust tier | The user-granted level of autonomy. New users start as observers; higher tiers do not bypass other policy checks. |
| Provenance | Where an action-relevant input came from. Missing provenance fails safe as untrusted external content. |
| Risk assessment | Mandatory evaluation of a candidate’s consequences and reasoning. Risk is not optional model metadata. |
| Spend limit | A hard per-action or daily financial bound. An amount above it must escalate. |
| Admission | The separate determination that an action is allowed to reach an execution adapter after policy and safety checks. |
Inference terms
| Term | Meaning |
|---|---|
| On-device / local inference | Reasoning performed through a locally available runtime and model. It does not silently fall through to a hosted provider. |
| Hosted provider | A configured remote model provider. Using one is an explicit network choice. |
| Verified-private inference | A provider-specific remote route admitted only after SkyTwin verifies live evidence for the exact call. TrustedRouter is available for explicit interactive calls that pass its pinned checks; NEAR AI remains verification-pending and unavailable. |
| Inference receipt | A structured record for a completed model-call attempt. It is not a claim that prompt content is encrypted or that a provider is confidential. |
Integration terms
| Term | Meaning |
|---|---|
| MCP | A local protocol surface for giving an agent scoped context and proposal capability. It is not universal delegation. |
| Connector | A component that can normalize a service’s information into signals. Source code for a connector is not a release-support claim. |
| Execution router | The component that selects from admitted adapters and provides a safety backstop; it does not replace policy. |
| DXT | A planned desktop extension transfer artifact for moving configuration between machines, not OAuth secrets. |