Data Science with Runtime-Native Execution

Documentation status: guide — see Maturity and evidence.

Use Runtime-native numerical execution when data science must operate close to conceptual memory, embeddings or model execution rather than exporting every workload to an external process.

Good fits

• feature extraction from conceptual structures;
• tensor projections connected to stable conceptual identity;
• graph/hypergraph neural workloads;
• embedding generation or search;
• inference where resident accelerator buffers reduce transfer overhead.

Keep boundaries explicit

A numerical result is not automatically a conceptual fact. Record how it was produced, its model/version and confidence where relevant, then use explicit rules or application policy to decide how it affects conceptual state.

Backend selection remains an execution concern; application semantics should survive a compatible CPU/GPU/backend change.