TR-Core is a symbolic cognitive architecture that autonomously discovers compositional rules from raw interaction without pretrained weights, embeddings, or LLMs. It achieves structural out-of-distribution generalization through an Apriori-style induction engine with exact binomial statistical filtering.
| Experiment | Task | TR-Core | Best Baseline | Ablation (no COMPOSE) |
|---|---|---|---|---|
| Phase 3 | Rare Compositional Rules (8 attrs, 4,608 contexts) | 0.78–0.80 balanced acc | 0.81 (ValueDListML) | 0.20 (chance) |
| Phase 4 | Structural-OOD (held-out triple, zero support) | 0.96–1.00 held-out acc | 0.98 (Decision Tree) | 0.00 |
Key Finding: COMPOSE operation is the causal mechanism for structural-OOD generalization — ablations collapse to chance level.
┌─────────────────────────────────────────────────────────────┐
│ ATTRIBUTE TASK ENVIRONMENT │
│ (Gymnasium: V3 Rare / V4 Structural-OOD) │
└─────────────────────────┬───────────────────────────────────┘
│ observe(context, action, success)
▼
┌─────────────────────────────────────────────────────────────┐
│ BEHAVIORAL PROGRAM V3 │
│ ┌─────────────────┐ ┌─────────────────────────────────┐ │
│ │ RULE BASE │ │ INDUCTION ENGINE │ │
│ │ • LATENT │ │ • Coverage pruning (c_min=4) │ │
│ │ • ACTIVE │ │ • Binomial filter (α=1e-3) │ │
│ │ • RETIRED │ │ • Precision ≥ 0.95 │ │
│ │ • DELETED │ │ • Max arity K=4 │ │
│ └────────┬────────┘ └──────────────┬──────────────────┘ │
│ │ │ │
│ └──────────┬───────────────┘ │
│ ▼ │
│ SUBPROGRAMS (COMPOSE) │
└─────────────────────────────────────────────────────────────┘
7 Development Operations: CREATE, GENERALIZE, SPECIALIZE, SPLIT, COMPOSE, MODIFY, DELETE
# Clone and install
git clone https://github.com/Mardon-programm/tabula-rasa-core.git
cd tabula-rasa-core
# Using uv (recommended)
uv venv .venv && source .venv/bin/activate
uv pip install -e .[dev]
# Or with pip
pip install -e .[dev]# Phase 3: Rare compositional rules
python -m experiments.run --experiment phase3 --seeds 1-24 --summarize --paired-stats
# Phase 4: Structural-OOD generalization
python -m experiments.run --experiment phase4 --seeds 1-24 --summarize --paired-stats
# Full DVC pipeline
dvc reprostreamlit run web/app.py
# Select "Phase 4 Demo" in sidebar → Initialize → Collect → Develop → Testimport gymnasium as gym
from environment.gymnasium_env import register_envs
register_envs()
env = gym.make("TRCore-AttributeTaskV4-v0", seed=42)
obs, info = env.reset()
action = env.action_space.sample()
obs, reward, terminated, truncated, info = env.step(action)
print(f"Success: {info['success']}, Oracle: {info['oracle_action']}")| Component | Tool |
|---|---|
| Configuration | Hydra (YAML, sweep-ready) |
| Pipeline | DVC (14 stages) |
| Statistics | Permutation tests (20k), Cohen's dz, bootstrap CI |
| Environments | Gymnasium (TRCore-AttributeTaskV3-v0, V4-v0) |
| Seeds | 24 seeds × 2 replications (main + replication) |
tabula-rasa-core/
├── behavioral/ # Core induction engine
│ ├── program.py # BehavioralProgram (v0.2)
│ ├── program_v3.py # Apriori induction (Phase 3/4)
│ └── rules.py # Rule, Condition, Literal, Evidence
├── environment/ # Task environments
│ ├── attribute_tasks_v3.py
│ ├── attribute_tasks_v4.py
│ └── gymnasium_env.py # Gymnasium interface
├── experiments/ # Experimentation framework
│ ├── experiment_phase3.py
│ ├── experiment_phase4.py
│ ├── analyze_results.py
│ ├── generate_figures.py
│ └── harness.py
├── conf/ # Hydra configs
├── docs/ # Paper, formal analysis
├── dvc.yaml # Reproducible pipeline
├── pyproject.toml # Package config
└── web/app.py # Streamlit demo
@article{trcore2026,
title={TR-Core: Autonomous Symbolic Induction of Compositional Rules via Apriori Search with Binomial Filtering},
author={TR-Core Research Group},
year={2026},
note={arXiv preprint}
}MIT License — see LICENSE for details.
- Alan — Next-gen embodied agent building on TR-Core: github.com/Mardon-programm/alan
- TR-Core Paper — docs/tr_core_paper.tex
- Formal Analysis — docs/apriori_formal_analysis.tex
Independent research by Mardon (School No. 7, Osh, Kyrgyzstan)