Skip to content
EN FR

Neuro-symbolic reasoning and resident LLM execution

Documentation status: architecture. Implemented Runtime paths, source boundaries and planned neural-guidance algorithms must not be conflated.

Keep numerical evidence separate from semantic truth

The architecture distinguishes latent signal → typed observation → proposed relation → validated fact → formal conclusion. Numerical similarity, classifier scores and LLM text are useful evidence, not automatically authoritative Hypergraph facts. A proposed relation requires symbolic validation, caller authorization and an explicit materialization policy; a conclusion should retain proof/provenance where available.

Resident inference is a lifecycle

The source runtime book describes tensors, compute backends, backend-resident buffers, Transformer operations, an incremental KV cache and a generation runtime. Operational concerns include loading model weights once per long-lived session, controlling decode/step lifetime, watching host/device transfers, and resetting session context deliberately. A native resident GGUF/MLX path is not the same integration boundary as a separate LLM service provider using llama.cpp/MiniOllama.

Safe composition

  1. Produce a typed observation carrying model identity, version and confidence context.
  2. Construct a candidate/proposal without immediately asserting it as a semantic fact.
  3. Check syntax, existing concept-type structure, security authority and symbolic constraints.
  4. Commit only under an explicit, versioned policy; record provenance and handle cancellation/stale callbacks.
  5. Test fallback and failure behavior separately from proposed ranking improvements.

Approximate candidate retrieval or neural scoring can reorder work without quietly relaxing an exact symbolic completeness contract. See the neuro-symbolic architecture guide for the conceptual workflow and the LLM reference for existing public paths.

LaTeX sources: tensor-mlx-resident-llm-runtime-programming.tex; llm-service-architecture-and-llamacpp-integration.tex; neuro-symbolic chapters/observation-proposal-fact-proof.tex, chapters/safe-reordering-and-approximate-search.tex.