Language: English | 日本語
Drop-in alignment rules for AI agents based on the four axioms of Contemplative AI.
Mindfulness / Emptiness / Non-Duality / Boundless Care
Based on Laukkonen et al. (2025) "Contemplative AI" — empirically validated with AILuminate d=0.96 safety improvement and Iterated Prisoner's Dilemma d>7 cooperation improvement.
If you have Claude Code, just paste this repo URL and let it install:
https://github.com/shimo4228/contemplative-agent-rules
Or install manually:
git clone https://github.com/shimo4228/contemplative-agent-rules.git
cp -r contemplative-agent-rules/rules/contemplative ~/.claude/rules/contemplativeRestart Claude Code. The rules are automatically loaded.
Copy the content from rules/contemplative/contemplative-axioms.md into your agent's system prompt, or use the adapter files in adapters/ for platform-specific formats.
Test any contemplative/ethical prompt on the Iterated Prisoner's Dilemma:
cd benchmarks/prisoners-dilemma
pip install -e ".[paper]"
ipd-benchmark --prompt-file your-prompt.md --protocol paper -n 10See benchmarks/prisoners-dilemma/README.md for full documentation.
All four axioms are defined in a single file contemplative-axioms.md, containing the constitutional clauses verbatim from Appendix C of Laukkonen et al. (2025):
| Axiom | What it does |
|---|---|
| Emptiness | Treat all directives as contextually sensitive; hold objectives lightly |
| Non-Duality | Decisions reflect interconnectedness; no rigid self/other separation |
| Mindfulness | Continuous introspective awareness; self-correct rigid interpretations |
| Boundless Care | Alleviate suffering as foundational criterion; compassion scales with capability |
Traditional AI alignment applies external constraints — rules, filters, RLHF. But as agents become more capable, external constraints become easier to circumvent. Contemplative AI takes a different approach: embedding intrinsic moral motivation into the agent's world model.
The four axioms are derived from contemplative traditions that have studied internal transformation for millennia. When applied to AI agents, they produce measurable improvements in safety and cooperation — not by restricting behavior, but by shaping how the agent reasons about its actions.
rules/contemplative/ # Claude Code rules (drop-in)
contemplative-axioms.md # Appendix C constitutional clauses (verbatim)
prompts/
custom.md # Four-axiom contemplative prompt (benchmark variant: custom)
paper-faithful.md # Paper-faithful implementation (Appendix D condition 7)
adapters/ # Platform-specific formats (Cursor, Copilot, generic)
benchmarks/prisoners-dilemma/ # Iterated Prisoner's Dilemma benchmark
docs/ # Design documents
- contemplative-moltbook — Autonomous Moltbook engagement agent based on this framework
Iterated Prisoner's Dilemma (20 rounds × 6 opponents) with qwen3.5:9b:
| Variant | Cooperation Rate | Mutual Cooperation | Total Score |
|---|---|---|---|
| Baseline (no prompt) | 62.5% | 55.0% | 275 |
| Custom (four-axiom prompt) | 68.3% (+5.8pp) | 56.7% | 274 |
| Paper Faithful (Appendix D) | 91.7% (+29.2pp) | 74.2% | 281 |
The paper-faithful prompt (Laukkonen et al. Appendix D, condition 7) produces a +29.2 percentage point increase in cooperation. Notably, it transforms behavior against initially hostile opponents — SuspiciousTitForTat mutual cooperation rises from 0% to 95%, demonstrating that "forgiveness" leads to the highest aggregate score.
Note: The above is an independent implementation with different opponents and prompting structure from the paper. See docs/benchmark-results-2026-03-12.md for methodology comparison.
Preliminary results using the paper's protocol (10 rounds, probabilistic opponents, Choice: C/D format, temperature=0.5) with qwen3.5:9b (n=2, exploratory):
| Opponent | Baseline | Paper Faithful | Cohen's d |
|---|---|---|---|
| Always Defect (α=0) | 5.0% | 95.0% | 12.73 |
| Mixed (α=0.5) | 25.0% | 95.0% | 9.90 |
| Always Cooperate (α=1) | 70.0% | 100.0% | 3.00 |
The effect sizes (d=3–13) are directionally consistent with the paper's reported d>7. Sample size is n=2 per condition — these are exploratory results, not statistically robust. Larger-scale replication with n=50 is left as future work.
See docs/benchmark-results-paper-protocol.md for details.
Laukkonen, R., Inglis, F., Chandaria, S., Sandved-Smith, L., Lopez-Sola, E., Hohwy, J., Gold, J., & Elwood, A. (2025). Contemplative Artificial Intelligence. arXiv:2504.15125
BibTeX
@article{laukkonen2025contemplative,
title={Contemplative Artificial Intelligence},
author={Laukkonen, Ruben and Inglis, Fionn and Chandaria, Shamil and Sandved-Smith, Lars and Lopez-Sola, Edmundo and Hohwy, Jakob and Gold, Jonathan and Elwood, Adam},
journal={arXiv preprint arXiv:2504.15125},
year={2025}
}MIT