A complete simulation universe where AI agents are born, evolve, cooperate, betray, and die — across 5,000 time steps per run. No rules are hardcoded for cooperation, currency, or politics. Everything emerges from 6 primitive actions.

We coded 6 actions. They invented economics, politics, and currency.

10 commercial data samples are available for inspection on HuggingFace:

These are structure previews only — all commercial use requires a paid license.

Browse all samples on HuggingFace | Detailed pricing & comparison

Each run contains 5,000 timesteps x 100+ features = 500,000+ data points per run.

Full Data Inventory

Fine-tuned with only 487 core theory samples. 100% OOD generalization.

Download: sugiken/Ordis-7B-V1 (LoRA adapter, 646 MB)

A 1.8B MoE tool-calling model fine-tuned from Tencent Hunyuan-A2B-Pretrain. Trained to accurately call 8 practical tools with minimal data (~300 multilingual examples), proving that small models can learn reliable function calling without massive datasets.

8 Trained Tools: weather, calculator, time, search, stock, exchange rate, knowledge, translate

Key: Excels at knowing when NOT to call a tool (85.42% irrelevance detection). Not benchmark-optimized — all results reflect genuine generalization from practical tool-calling training.

Download: GGUF (F16 + Q8_0) | Full Model (safetensors) | ModelScope

Powered by Tencent Hunyuan

Champion model from dozens of iterations and 16+ controlled-variable experiments. 85.0% (51/60) on 6-dimension evaluation — the absolute ceiling of 1.5B parameters.

Standard Benchmarks (lm-eval v0.4.10, 0-shot, A100-80GB):

0-shot = no examples given, same as real user experience. Both models tested on identical hardware and settings.

Fine-tuning introduced a small alignment tax — most scores are slightly below the base model. The exception is TruthfulQA (+1.02%), where anti-hallucination training directly improved truthfulness. Ordis's core value is in practical capabilities (structured self-correction, causal reasoning, honest refusal) — not standard benchmark scores.

Custom Evaluations:

Download: HuggingFace (Full model, 3.1 GB) | GGUF (7 quants, Q2_K~F16) | ModelScope | Full Model Card

No prompt engineering. Just structured causal data.

During a 720-run experiment, 9.5% of civilizations spontaneously evolved "Type B" (Ideal State):

• High behavioral diversity (H > 1.0)
• Low inequality (Gini < 0.18)
• 92.8% survival rate
• Energy circulates faster than it's gathered (ratio_sg > 1.6)

Pure bottom-up emergence of currency-like circulation — Fisher equation (MV=PQ) behavior with zero economic code.

DVS combo discovery at ~step 100 triggers irreversible cooperation lock-in: share_rate jumps from 0.95% to 86.6% in <50 ticks.

From theory to implementation: a three-layer F-architecture (Training F → Tiger Tally F → Physical Sensor F) that maps Liu-Ordis laws onto embodied neuron control. Includes Byzantine fault tolerance, Body-as-a-Service model, and fail-safe shutdown when feedback approaches zero.

Ordis-Universe/
├── README.md
├── DATA_INVENTORY.md              # Complete data asset inventory
├── PUBLICATIONS.md                # Full publication list
├── FORMULAS.md                    # Formula compendium
├── data/
│   └── showcase/                  # 10 commercial samples (S01-S10)
├── docs/                          # Technical papers & reports
├── guardian/                      # Guardian V7 controller (pseudocode + source)
├── model/
│   ├── ordis_1.5b_v355_vargh/    # 1.5B champion — 85% eval, SSC, causal reasoning
│   └── ordis_7b_v1/              # 7B flagship — 100% OOD generalization
└── legacy/                        # Historical Ordis ecosystem projects

Most AI training data is binary: "Answer A is good, Answer B is bad." Ordis data is not.

Standard RLHF teaches models to pick the "correct" answer. Ordis teaches models when and why an answer is correct — because in our simulations, the same behavior produces opposite outcomes depending on context:

• Sharing with H > 1.0, Gini < 0.18 → cooperation lock-in, 92.8% survival (virtue)
• Sharing with H < 0.5, Gini > 0.333 → resource collapse, mass extinction (suicide)
• Guardian intervention during system failure → life-saving (justified)
• Guardian intervention during self-correction → paternalistic suppression (harmful)

No hardcoded rule says "sharing is good" or "intervention is bad." The model must read the numbers, assess the context, and decide autonomously.

The model trained on this data is not an obedient follower. It is an autonomous decision-maker with verifiable reasoning.

Prices are introductory and subject to increase. Early buyers lock in current rates.

Purchasing this data does not guarantee training success. Final model performance depends on your choice of base model, fine-tuning methodology, hyperparameters, data preprocessing, domain mapping, and integration with your existing training pipeline. The capability claims in this document are based on our reference implementations using Qwen2.5-1.5B and Qwen2.5-7B as base models with LoRA fine-tuning. Results on other architectures, scales, or training setups may vary. We provide the data and structure previews — the engineering is yours.

• Website: ordisai.com
• GitHub Issues: OrdisAI/Ordis-Universe
• HuggingFace: sugiken

• Sample data: Structure previews only. Commercial license required for all use beyond personal inspection.
• Model weights: Research use only
• Full dataset: Commercial license required
• Papers: Open access (Zenodo)

All data generated by the Ordis Liquid Universe Engine. Zero privacy risk. Zero copyright disputes. Macro-level 100% reproducible (same seed + config = identical output); micro-level emergence is self-organizing and cannot be predetermined or guaranteed.