Symbolic_Reasoning

Symbolic Reasoning

Orrin reasons with explicit, inspectable representations — world models, causal graphs, and a full rule lifecycle — entirely without an LLM. This is the core of the symbolic-first design: the decision loop and world understanding are deterministic and auditable, not hidden inside a model's weights.

What it does

• World model (inference.py) — description-logic-style inheritance and queries over entities and predicates.
• Causal reasoning (causal_graph.py) — Pearl-style causal edges, interventions, and counterfactuals. The causal graph is a self-model, so frontier questions point inward (search Orrin's own code and state), not at generic web research.
• Prediction (prediction_engine.py) — predictive-processing; prediction errors feed control signals.
• Concepts and analogy (concept_formation.py, analogy_engine.py) — forms new concepts and maps analogies across domains.

The rule lifecycle

Rules are born, generalized, verified, and forgotten — a small self-maintaining knowledge base:

flowchart LR
    OBS[Observed regularity] --> SYN[Synthesize rule]
    SYN --> VER[Verify]
    VER --> ENG[Rule engine fires]
    ENG --> ABS[Abstract / compress]
    ABS --> FOR[Forget if it stops paying]
    FOR -.-> ENG

Loading

Autonomy comes from autonomous_experiment.py (designs and runs experiments to test its own hypotheses) and the learning-progress signals in intrinsic_motivation.py / progress_tracker.py. Productions — synthesized principles, crystallized skills, resolved experiments, causal edges — are recorded on the effect ledger so symbolic work earns reward the same way file output does.

For the full module map and how to extend it, see Symbolic Reasoning Subsystem and Extending Symbolic Operations.

Code pointers

• brain/symbolic/inference.py, causal_graph.py, prediction_engine.py
• brain/symbolic/rule_engine.py and the rule_*.py lifecycle modules