Skip to content
EN FR

Tensors — Numerical Execution Connected to Concepts

Documentation status: architecture — see Maturity and evidence.

The tensor subsystem provides numerical computation that can be connected to conceptual data and Runtime execution. It supports ordinary dense computation, conceptual/sparse projections, neural-network layers and accelerator-backed execution.

Architectural rule

Tensor state is a numerical projection. Conceptual identity remains owned by the conceptual model and hypergraph.

conceptual structure
  -> explicit numerical projection
  -> tensor computation
  -> numerical result
  -> optional reattachment to conceptual identity

Current capability areas

  • dense and conceptual tensors;
  • element-wise operations, reductions and linear algebra;
  • feed-forward and graph neural-network computation;
  • transformer-oriented execution;
  • capability-driven compute backends;
  • resident accelerator buffers;
  • WebGPU execution paths;
  • quantized weight execution.

Reference