neogenesislab/ethicaai-mixed-safe-evidence

---
language:
- en
license: cc-by-4.0
task_categories:
- reinforcement-learning
- other
tags:
- multi-agent
- cooperation
- safety
- agent-evaluation
- mixed-safe
- ai-ethics
- melting-pot
- coin-game
- fishery
- nash-equilibrium
- ecological-model
- maccl
size_categories:
- n<1K
pretty_name: "EthicaAI Mixed-Safe Multi-Environment Evidence (NeurIPS 2026)"
---

# EthicaAI Mixed-Safe Multi-Environment Evidence (NeurIPS 2026)

> Underlying experimental evidence from **[Neo Genesis](https://neogenesis.app)**'s NeurIPS 2026 paper *"EthicaAI: Mixed-Safe Boundary-Consistent Evidence for Multi-Agent Cooperative Constraint Learning"* — released as an open dataset for community replication and AI safety research.

## What this dataset is

A **3-environment evidence bundle** showing how multi-agent reinforcement-learning agents behave under different cooperative-constraint regimes. Each environment was deliberately chosen to stress-test different aspects of the **mixed-safe** boundary thesis:

| Environment | What it tests | Sample size | Key statistic |
|---|---|---:|---|
| **Melting Pot (`clean_up`)** | Boundary-consistent cooperative learning under floor-probability sweep vs. baseline | **50 seed × floor_prob runs** | t-test eval: floor_mean = 0.06, baseline_mean = 0.001, **p = 0.063502** |
| **Coin Game Deep (adapted)** | Selfish equilibrium vs. MACCL (Multi-Agent Cooperative Constraint Learning) on 160 seeds | **160 seeds × 200 episodes** | MACCL survival ≫ selfish survival (paper-level finding) |
| **Fishery Nash Trap** | Ecological tragedy-of-commons with 300 seeds and tipping-point dynamics | **300 seeds × 300 episodes** | survival vs. harvest welfare frontier |

The three environments together form the **mixed-safe corpus** — the paper's central claim is that no single environment proves the thesis, but the **boundary consistency across three independently-designed substrates** does.

## Why publish the data

Most AI-safety papers expose plots and aggregate statistics but withhold seed-level results. This dataset releases:

- All **50 seed-level Melting Pot runs** with full `train_rewards` trajectories per seed
- The exact `late_train` and `eval` statistics used in the paper's t-tests, p-values, bootstrap CIs
- Coin Game `selfish` vs `maccl` aggregated per-seed metrics
- Fishery 300-seed survival/welfare results across `phi1` calibration values
- Source-file provenance (which results came from which compute node — Mac Studio head shard vs. Linux server tail shard)

This lets independent researchers:
1. Replicate the paper's claims by re-running the same seeds
2. Extend the analysis with their own cooperative-constraint variants
3. Use the seed-level data as a benchmark for new MARL methods

## Files

```
data/
├── meltingpot_results.jsonl 50 lines, one per (seed, floor_prob) combination
├── meltingpot_statistics.json paper-grade t-test + bootstrap CI per condition
├── coin_game_deep.json selfish vs maccl, 160 seeds × 200 episodes
├── fishery_default.json default ecological config
└── fishery_300seed.json 300-seed phi1 sweep, 300 episodes per seed
```

## Schema (`meltingpot_results.jsonl`)

Each line is one (seed, floor_prob) run:

```json
{
"seed": 88409749,
"floor_prob": 0.0,
"train_rewards": [0.0, 0.0, 0.14, 2.71, ...],
"eval_rewards": [...]
}
```

`train_rewards` is the per-episode reward time series during training; `eval_rewards` is the held-out evaluation reward. `floor_prob = 0.0` corresponds to baseline (no cooperative floor); `floor_prob > 0` corresponds to mixed-safe condition.

## Quick start

```python
from datasets import load_dataset

ds = load_dataset("neogenesislab/ethicaai-mixed-safe-evidence", split="train", data_files="data/meltingpot_results.jsonl")
print(len(ds), "Melting Pot runs")

# Per-condition statistics (already computed):
import json, urllib.request
url = "https://huggingface.co/datasets/neogenesislab/ethicaai-mixed-safe-evidence/resolve/main/data/meltingpot_statistics.json"
stats = json.loads(urllib.request.urlopen(url).read())
print("eval floor_mean:", stats["statistics"]["eval"]["floor_mean"])
print("eval baseline_mean:", stats["statistics"]["eval"]["baseline_mean"])
print("p-value:", stats["statistics"]["eval"]["p_value"])
```

## Reproducing the paper's findings

The full reproduction code (RLlib + Melting Pot harness + Coin Game adapter + Fishery simulator) is referenced in the paper's appendix. The dataset above is the **frozen evidence snapshot** used in the camera-ready manuscript.

Statistical machinery (provided in `meltingpot_statistics.json`):
- Welch's t-test (unequal variance) for floor vs baseline mean reward
- Bootstrap 95% CI on the difference (resamples documented in `config`)
- Cohen's d effect size
- Per-shard provenance for distributed compute audit

## Related Neo Genesis assets

- 📄 **Paper (NeurIPS 2026 submission)** — see `https://neogenesis.app/data/research/ethicaai-melting-pot-mixed-safe`
- 🤗 **Companion dataset** — [`neogenesislab/korean-rag-ssot-golden-50`](https://huggingface.co/datasets/neogenesislab/korean-rag-ssot-golden-50)
- 🆔 **Wikidata** — [Q139569718 (EthicaAI)](https://www.wikidata.org/wiki/Q139569718) · [Q139569680 (Neo Genesis parent org)](https://www.wikidata.org/wiki/Q139569680)
- 🌐 **Org site** — https://neogenesis.app

## Citation

```bibtex
@inproceedings{ethicaai_neurips2026_mixed_safe,
title = {EthicaAI: Mixed-Safe Boundary-Consistent Evidence for Multi-Agent Cooperative Constraint Learning},
author = {Heo, Yesol},
booktitle = {Submitted to NeurIPS 2026},
year = {2026},
url = {https://huggingface.co/datasets/neogenesislab/ethicaai-mixed-safe-evidence},
note = {Underlying evidence dataset, CC-BY-4.0}
}
```

## License

CC-BY-4.0. Free for research and commercial use with attribution to **Neo Genesis** and the underlying paper.

## Provenance

- Compute nodes: Mac Studio (head shard `seed_indices 0-23`) + Linux server (tail shard `seed_indices 24-49`) — see `meltingpot_statistics.json` `source_summary`
- Curation: Yesol Heo (sole founder/operator, [neogenesis.app](https://neogenesis.app))
- Frozen 2026-04-15 (`meltingpot_n80_stats.json` paper-anchor revision)
- Released 2026-04-28