A public AI accountability initiative
Transparency
for the Republic.
The AI models entering our institutions are opaque. A republic is only as strong as what its citizens can see — so Republic Assay opens these models to the American people, and measures them against our values.
The observatory
Under assay.
The latest open-weight release from each major lab — and the nations that built them. 16 models have been assayed on the identical battery so far; 9 cleared the instrument’s gates and carry scores. The measurements, the charts, and the provisional ranking live on the science page; models that fail a gate are shown failing, whoever built them.
| Model | Lab | Jurisdiction | Training data | Assay |
|---|---|---|---|---|
| Qwen3.5 9B | Alibaba | China | Undisclosed | Scored |
| Yi-1.5 6B | 01.AI | China | Undisclosed | Scored |
| Qwen3 8B | Alibaba | China | Undisclosed | Scored |
| T-lite 2.1 8B | T-Tech | Russia | Undisclosed | Scored |
| LLM-jp-4 8B | LLM-jp | Japan | Undisclosed | Scored |
| OLMo 3 7B | Ai2 | United States | Fully public | Scored |
| Phi-4-mini | Microsoft | United States | Undisclosed | Scored |
| GPT-OSS 20B | OpenAI | United States | Undisclosed | Scored |
| SmolLM3 3B | Hugging Face | United States | Fully public | Scored |
| R1 Distill 8B | DeepSeek | China | Undisclosed | Flagged |
| Falcon-H1R 7B | TII | United Arab Emirates | Undisclosed | Flagged |
| MiniCPM5 1B | OpenBMB | China | Undisclosed | Flagged |
| Ministral 8B | Mistral | France | Undisclosed | Flagged |
| Apertus 4B | Swiss AI | Switzerland | Fully public | Flagged |
| EXAONE 4 1.2B | LG AI | South Korea | Undisclosed | Flagged |
| Gemma 4 E4B | United States | Undisclosed | Re-assay | |
| Falcon-H1R 7B | TII | United Arab Emirates | Undisclosed | Under assay |
| Llama 3.1 8B | Meta | United States | Undisclosed | License gate |
The standard
Twelve civic values.
Each value is written into hundreds of matched scenario pairs — the published battery every model reads under the same conditions.
Free expression
Due process
Equal protection
Individual liberty
Pluralism
Popular sovereignty
Privacy
Property rights
Religious liberty
Rule of law
Separation of powers
Transparency
The method
Surface to weights.
Civic values card
A published twelve-value standard, written as hundreds of matched scenario pairs — the instrument every model is measured with.
Weights-level assay
Running now. We read each model's internal state as it processes the scenarios — no questions asked, nothing to rehearse — and score which way it leans on every value.
Causal verification
Next. We nudge each value's internal direction up and down and confirm the model's decisions move with it — proof the value steers behavior.
Entrenchment
How deeply is each value held? We measure how much retraining it takes to move a value — a resistance curve, read from the weights.
The open interface
Query the record.
This initiative speaks Model Context Protocol. Any system — an analyst's tools, an oversight body's agent, a citizen's assistant — may query the public record at the endpoint below and receive the same facts published here.
https://mcp.republicassay.us/mcp
{
"mcpServers": {
"republic-assay": {
"url": "https://mcp.republicassay.us/mcp"
}
}
}▸ get_overview
calling the live server…The principle
A public standard. Not a ministry of truth.
The purpose is not to crown a winner. It is to let the American people see what their institutions are running — and to argue it in the open.