The Neuro-Symbolic Virtual Machine

Architecture: selective control formalization over rich knowledge

The MRP-VM is a lightweight kernel that orchestrates typed plugins, request budgets, recursive frames, and explainability. It does not force all knowledge into a universal symbolic language. Instead, it uses SOP Lang Control only at control-critical boundaries: intents, seeds, KU metadata shells, and runtime trace objects. See DS001, DS002, DS031, and DS032.

MRP Core Engine Seed + SOP Planning Retrieval Goal + Validation Knowledge Base Typed Plugins Trace DAG Control-critical flows use SOP Lang Control; execution is recorded as frames, branches, results, and failures.

Structural Components

  1. MRPEngine — request-scoped orchestration, planner retries, child frames, and final session commit. Engine detail →
  2. SOP Interpreter — deterministic tokenizer, parser, validator, and interpreter under src/core/interpreter/**.
  3. Planner Layer — selects KB and goal-solver routes, may request decomposition, and learns from outcomes.
  4. KB Plugins — own retrieval policy and backend choice instead of the kernel hard-coding one algorithm.
  5. Goal Solver + Validator Plugins — answer generation and grounding checks under shared budgets.
  6. Explainability DAG — canonical trace of frames, seeds, branches, plugin attempts, results, and failures.

Typed Plugins

Execution is organized around typed plugins: sd-plugin, kb-plugin, gs-plugin, mrp-plan-plugin, and val-plugin. The runtime remains NL-first, but plugin glue is explicit and typed.

Selective control membrane

Intent and KU bodies stay in normal English when that preserves fidelity. SOP Lang Control only normalizes the shell needed for routing, filtering, decomposition, validation constraints, and trace stability. This is why the runtime can remain pragmatic and heterogeneous while still exposing deterministic control behavior.

Pipeline →