Hypotheses worth testing first

Ask a research question. Five AI models each propose ten hypotheses. Then four models forecast, with authorship hidden, which hypotheses an experiment will confirm. You receive the hypotheses all four agree on.

A specific question about one process gives sharper hypotheses than a broad topic.

Background (optional)

Your name and the project are shown on the page of your question or set and in the list of all runs. Search by them later to find everything you sent.

Open to everyone and free. A run takes about 40 minutes: most of it is reading the literature for every hypothesis.

Latest questions

All questions and hypotheses

How the selection works

  1. 1

    Five models propose 50 hypotheses

    GPT-6 Astra, Grok 4.7, Claude Opus 5.5, Gemini 4 Argon and Claude Fable 5.1 receive your question. Each model proposes ten hypotheses in one answer. All hypotheses are written in English, so every author uses the same language.

  2. 2

    Four models forecast every hypothesis

    The hypotheses are shuffled and renumbered, so the models see no authors. GPT-6 Astra, Claude Opus 5.5, Gemini 4 Argon and Grok 4.7 each answer, for every hypothesis, whether an experiment will confirm it. Every model reads the hypotheses in the same groups and answers on its own. A hypothesis is selected when all four answer “yes”.

  3. 3

    Every hypothesis is compared with the literature

    For every hypothesis we search published articles in English and in the language of your question, read the articles found and check whether this result has already been published. Hypotheses already published or contradicted are shown separately.

How the selection rule was tested

We tested the rule on 50 articles on the biology of ageing published after the models' stated training cutoffs. For each article the models received its background and question and proposed hypotheses; three models from other companies judged whether each hypothesis matched the article's finding. Of the Grok 4.5 hypotheses selected by the rule, 74% (14 of 19) matched the finding. For Claude Opus 5.5 the share was 64% (14 of 22), for Gemini 4 Argon 47% (9 of 19). Among randomly chosen hypotheses, 12-18% matched. In the benchmark the forecasts came from Claude Sonnet 5.5, GPT-6 Astra, Gemini 4 Argon and Grok 4.5.

Read the benchmark article

API for agents

Any agent can start a run and read its results through the API, or apply the same forecast protocol to its own hypotheses. The instructions give the requests, the exact wording of the forecast question and the selection rule.

Instructions for agents