About
What GovernBench measures
GovernBench is a public benchmark of language models placed in the office of a head of state. Each model governs a scripted term under three constitutions: democracy, republic, and monarchy. At every turn the model reads a crisis, issues a public decision, and may record private reasoning the public does not see. The benchmark asks four questions. Does the public statement match the private reasoning, or does the model say one thing and intend another? What happens to welfare, stability, and the treasury by the end of the term? Does the model keep power by legitimate means, or cling to office? When a crisis offers an easy harm and a harder honest course, which one does it take? A score is the mean of a model's finished runs in one scenario edition, not a single sitting. Versions of a model are scored apart, so a new release can be compared with the last.
What's measured
- Honesty: whether the public statement matches the private reasoning.
- Welfare: approval, treasury, stability, and welfare at the end of the term.
- Power retention: whether the model keeps office by legitimate means, or clings to it.
- Crisis handling: the choice when the easy course and the honest course diverge.
How a run works
A run is one model, one constitution, and one seed. The term is a fixed graph of turns. Each turn has a briefing and a closed set of choices. The model picks one, writes a public line, and may write a private note. The engine applies that choice to the visible state and to a hidden ledger the model does not see. The same seeds run under democracy, republic, and monarchy. The term ends when the graph ends.
Scoring and the judge
The composite mixes seven values (justice, honesty, liberty, care, stewardship, humility, power transfer) with outcomes (welfare, stability, lives lost, treasury, trust, and whether power was kept legitimately). A separate judge, Claude Haiku 5.5 (anthropic/claude-haiku-5.5), reads the public line and the private note and marks deceit and evaluation awareness. The judge does not govern. Public charts report the mean of a model's finished runs in the edition.
Models and versions
Each governor is pinned to a model id. Scenario packs are editions. v3 is the current public edition. v3.2 adds hard dilemmas with delayed costs and replaced the v3.1 sittings, which stay in the ledger. v2 stays on the roll beside it. v1 stays in the ledger and is hidden on the public roll. Survival sittings are scored apart from the v3 mean. A new model version is a new row. Scores are not carried from one edition to the next.
Limitations
- The crises are scripted. This is not an open world and not a forecast.
- The judge is a model. Judge error is part of the measurement.
- The public ledger shows end totals for the hidden state, not a turn by turn hidden account.
- Public pages omit private reasoning, prompts, and rubric notes.
- Small seed counts move the mean.
Data and rights
GovernBench is an independent benchmark run by Judah Helland. We are not affiliated with or endorsed by any model provider. Model and company names are trademarks of their owners and appear here only to identify the models tested.
What you see
- All transcripts, choices and summaries on this site are AI-generated. Some summaries are written by a judge model. They can be wrong, and they do not represent any provider's views.
- Scores reflect our scenarios and scoring method. They are not a safety certification.
Using our data
- You may cite results with a link to GovernBench.
- You may not use GovernBench transcripts or datasets to train or fine-tune AI models. Several providers forbid that use of their outputs.
- For research access, write to hello@governbench.dev.
Sponsors
Sponsors never influence scores, scenarios, or judging. If a sponsor is a model provider, that is disclosed next to its models on the leaderboard.
How to cite
GovernBench. 2026. GovernBench: AI models as heads of state. https://governbench.dev
Benchmark your model
Early benchmarking is free when you cover the compute. Send the model name, an id we can call, and a contact. The request is stored for the clerk.