Skip to content
EN FR

Neuro-symbolic architecture: evidence and proof

Documentation status: architecture. This page summarizes an architecture working draft. Planned ranking, approximate search or learned guidance is not automatically a current supported application API.

Five layers of authority

MODEL owns editable declarations; APPLICATION assembled metadata; SEMANTIC symbolic concepts, relations, identities and proof provenance; RUNTIME execution state; NUMERICAL embeddings, vector indexes, tensors and ranking statistics. Numerical structures accelerate work but remain derived.

Decision path for one candidate

Neural / latent signal
   -> typed observation (source + confidence + provenance)
   -> proposed relation (not a fact)
   -> symbolic validation / authorization
   -> governed fact materialization
   -> formal conclusion with proof trace when applicable

Do not treat similar vectors as equivalent concepts. A candidate rejected by a symbolic type constraint must not be revived by a strong numerical score. Safe ranking can change candidate order, not validity. If approximate pruning changes completeness, the contract must declare it; exhaustion of a resource budget is not logical refutation.

Reduce semantic work before micro-optimization

Favor canonical identities, selective indexes, delta updates and explicit fallback. If a derived cache becomes stale, rebuild, bypass or fall back to the exact path when the contract promises exactness. A neural component is optional unless the deployed capability contract says otherwise.

What to check in a release

Distinguish the parser and hypergraph capabilities actually tested from advanced architectural scenarios. Verify score-provider contracts, cancellation/stale callbacks, inference provenance, security and semantic-conformance tests before presenting a new neuro-symbolic optimization as delivered.

Continue

Hypergraph/H-Logic programming · Native LLM execution · Knowledge engineering · Evidence policy

LaTeX sources: logicells-neuro-symbolic-architecture-guide/chapters/core-principles.tex, authority-layers-and-representations.tex, observation-proposal-fact-proof.tex, safe-reordering-and-approximate-search.tex.