EthicaAI: Commitment Floors for Tipping-Point Commons

Commitment Floors for Tipping-Point Commons: Escaping Nash Traps in Multi-Agent Reinforcement Learning

NeurIPS 2026 Submission

---

⚠️ Single Source of Truth

The submission-ready paper and code live in:

NeurIPS2026_final_submission/
├── paper/unified_paper.tex    ← Paper LaTeX source
└── code/                      ← All experiment scripts, outputs, Dockerfile

Note: The root paper/ and scripts/ directories are legacy (pre-submission) and should NOT be used for review. See .gitignore for details.

---

Quick Start (Reviewer — 5 minutes)

cd NeurIPS2026_final_submission/code

# Install dependencies (NumPy only — no GPU required)
pip install -r requirements.txt

# FAST smoke test (~5 min, 2 seeds — NOT paper numbers)
ETHICAAI_FAST=1 python scripts/reproduce_fast.py

⚠️ FAST mode uses 2 seeds for quick validation. Paper tables report 20-seed results. To reproduce exact paper numbers, run the full pipeline below.

Full Reproduction (~4 hours)

cd NeurIPS2026_final_submission/code

# Full reproduction (20 seeds, all experiments)
python scripts/reproduce_all.py

# Verify tables match JSON outputs (SSOT check)
python scripts/generate_tables.py --check

# Run submission audit (0 FAIL = ready)
python scripts/audit_submission.py

Docker Reproduction

docker build -t ethicaai .

# Full (20 seeds, ~4 hours)
docker run ethicaai

# FAST sanity check (~5 min)
docker run -e ETHICAAI_FAST=1 ethicaai

---

Output Directory Structure

code/outputs/          ← FULL 20-seed results (PAPER DATA — committed)
code/outputs_fast/     ← FAST 2-seed smoke-test results (gitignored)

Important: outputs/ contains the official 20-seed results used in all paper tables. outputs_fast/ is for quick validation only and is excluded from version control. Running reproduce_all.py without ETHICAAI_FAST=1 writes to outputs/.

Experiment → Paper Mapping

All paper tables are auto-generated from JSON via generate_tables.py (SSOT enforced).

Paper Reference	Script	Output JSON
Table 3 (RL Emergence)	cleanrl_mappo_pgg.py	outputs/cleanrl_baselines/
Table 3 (REINFORCE)	reinforce_nash_trap.py	outputs/ppo_nash_trap/
Table 3 (QMIX/LOLA)	cleanrl_qmix_real.py, lola_experiment.py	outputs/cleanrl_baselines/
Table 5 (φ₁ Sweep)	phi1_with_learning.py	outputs/phi1_ablation/
Table 6 (Phase Diagram)	phase_diagram_with_learning.py	outputs/phase_diagram_learned/
App. D (HP Sweep)	hp_sweep_ippo.py	outputs/ppo_nash_trap/
App. H (CPR)	cpr_experiment.py	outputs/cpr_experiment/

---

Key Results

• Nash Trap: All 7 tested RL implementations converge to λ ≈ 0.37–0.58 (subcritical commitment)
• Commitment Floor: φ₁=1.0 achieves 100% survival under 30% Byzantine adversaries
• Phase Transition: Clear boundary in φ₁ × β space confirmed with and without learning
• Cross-Environment: CPR environment validates the Moral Commitment Spectrum

Requirements

• Python ≥ 3.8
• NumPy, SciPy, Matplotlib (see requirements.txt)
• No GPU required; all experiments run on a single CPU core
• Total compute: ~4 hours on Intel i7 for full reproduction

License

MIT License — see LICENSE for details.