v2 · February 2026 · NeurIPS Submission

EthicaAI
Beyond Homo Economicus

Computational Verification of Sen's Meta-Ranking Theory
via Multi-Agent Reinforcement Learning

Yesol Heo · Neo Genesis Research

Explore 70 Figures ↓ Zenodo DOI ↗ GitHub ↗

At a Glance

560+ experiments. 70 figures.

0
Experiment Runs
0
Figures
0
Environments
0
Research Stages
0
Max Agents
p=.0023
HAC Robust SE
EthicaAI formalizes Amartya Sen's "Meta-Ranking" theory — preferences over preferences — as a dynamic mechanism in Multi-Agent Reinforcement Learning. We demonstrate that Situational Commitment — morality conditional on survival — is the only Evolutionarily Stable Strategy across 4 environments, 7 SVO conditions, and up to 1,000 agents.

Key Findings

What we discovered

1

Dynamic meta-ranking enhances collective welfare significantly

p=0.0003 · Cohen's f²=0.40
2

Agents exhibit emergent role specialization — Cleaners vs Eaters

σ divergence p<0.0001
3

Only "Situational Commitment" survives as Evolutionarily Stable Strategy

~12% of population · 200-gen
4

Individualist SVO best matches human PGG behavioral data

Wasserstein Distance = 0.053
5

SVO rotation accounts for majority of total causal effect

86% · full factorial 2³
6

Human-AI behavioral alignment maintained across all conditions

WD < 0.2 all conditions
7

Byzantine robustness maintained with up to 50% adversarial agents

Up to 50% adversaries
8

Scale invariance verified from small groups to large populations

20 → 1,000 agents · ATE ±0.03

Interactive Playground

Try the simulation

Adjust SVO angle and toggle Meta-Ranking to see how agent cooperation changes in real time

Selfish 0° Prosocial 45° Altruist 90°
METRICS
Wasserstein D
—
Avg Contribution
—
Final λ
—
Contribution Rate
● Meta-Ranking ● Baseline ● Human

Research Journey

Seven stages

Stage 1: Core PGG
7 SVO conditions, causal inference framework, baseline experiments
Stage 2: Extended
Mixed-SVO populations, communication channels, continuous action spaces
Stage 3: Cross-Environment
Validation across IPD, PGG, Harvest — evolutionary dynamics
Stage 4: Human Alignment
Behavioral alignment with human PGG data, 100-agent scalability
Stage 5: MAPPO
Partial observability, LLM comparison, multi-resource environments
Stage 6: Applications
Vaccine allocation, AI governance, human-AI interaction scenarios
Stage 7: Robustness Final
Byzantine robustness, Moran process, GNN architecture, policy implications

Cite This Work

Use in your research

@article{heo2026ethicaai, title={EthicaAI: Beyond Homo Economicus -- Computational Verification of Sen's Meta-Ranking Theory via Multi-Agent RL}, author={Heo, Yesol}, year={2026}, url={https://ethica.neogenesis.app}, doi={10.5281/zenodo.18637742} }
Zenodo DOI ↗ GitHub ↗