TR-Core (Tabula Rasa Core) is an experimental symbolic R&D platform designed to study compositional generalization and structural out-of-distribution (OOD) reasoning. Unlike traditional deep neural networks and LLMs, the system operates without pre-trained weights and builds an interpretable world model through explicit rule induction.
The v0.3 architecture relies on three key components: dynamic Apriori rule induction (program_v3.py), deterministic evidence tracking for hypotheses, and a dedicated COMPOSE mechanism for synthesizing complex behavioral programs from atomic logical predicates.
The technical implementation is built in pure Python without heavy machine learning frameworks, featuring Gymnasium integration (TRCore-AttributeTaskV4-v0), a fully reproducible DVC pipeline, and rigorous statistical validation via permutation testing and bootstrap confidence intervals.
Key Results & Findings:
Phase 3 (Rare Compositional Rules, 8 attributes / 4608 contexts): Achieved 0.78–0.80 balanced accuracy ($p < 10^{-4}$), outperforming kNN (0.70) and Random Forest (0.66).
Phase 4 (Structural-OOD): Reached 0.958–1.000 accuracy on zero-train-support held-out configurations, whereas standard baselines degraded to random chance (~0.25).
Ablation Analysis: Disabling the COMPOSE module dropped accuracy to zero, experimentally confirming that explicit symbolic composition is the core mechanism driving structural transfer.
