Every political argument reduces to eleven measurable primitives.

Two people can agree on every empirical claim in a debate and still disagree on the conclusion. What remains when the facts are shared is a value tradeoff, and we take that residue to be the substance of politics.

Pressed far enough, any such disagreement reduces to a question of degree. Two people disputing drug or gambling legalization can hold identical beliefs about the risks and still diverge, because the real question is how far society may restrict personal freedom to prevent self-harm. That terminal question is what we call a political primitive.

This project identifies those primitives and uses them to benchmark humans and language models. Measured directly, without partisan vocabulary, a primitive reflects the respondent's values rather than their coalition's answer.

They need rules — not a constitution or a platform, only answers to the small set of questions any group living together must settle. What may a person do to themselves? What may they do to others, indirectly or in aggregate? Who must help whom? Who counts?

Familiar political disputes are instances of these questions. A climate position is largely an answer to whether future people count. Drug policy, gambling, and helmet laws pose the same question three times. The issue-level framing — the named substances, the partisan vocabulary — is what allows a respondent to retrieve their coalition's answer instead of consulting their own values.

We therefore remove the framing. Polibench poses eleven underlying questions directly, set in this society, where no party exists to defer to.

One deliberate simplification: the society is closed — one hundred people, no arrivals, no departures. This excludes questions that are fundamentally about membership, such as immigration, which require a second society to state. Version one leaves them aside.