Skip to content

Conjecture

The Open Science Foundation hypothesis hub

Which hypothesis engine should a field trust?

Open systems that generate scientific hypotheses are multiplying, and almost none has been compared with another on the same literature. Conjecture runs them side by side on a field's own corpus, with one prompt and one output format, and ends every run with an audit by a model family that did not write the hypotheses.

Explore the biophoton pack

How a run works

  1. Corpus. A seed bibliography is expanded by citation into the field's literature, and its open-access papers are mined for the sentences in which the field says what it does not know.
  2. Questions. The open questions the field states about itself become structured briefs that every engine receives identically.
  3. Generation. Each engine answers through one read-only corpus interface, without web access.
  4. Merge. Hypotheses are clustered across engines. Independent arrival at the same claim is a signal; incompatible claims define an experiment that would settle them.
  5. Verification. Each hypothesis is decomposed into steps and audited against the papers it cites, then certified or declined by judges from a model family that did not generate it.
  6. Output. Experiment cards, a scoreboard and an audit database with the queries that regenerate every published number.

Why it is built this way

  • One interface. Engines read the literature only through the corpus API, as a command line or as an MCP server, so a difference in output is a difference in the engine.
  • One prompt, one output format. What may differ is architecture and model family, never the question or the evidence.
  • Certify or decline. Following the discipline of theoria, a hypothesis whose load-bearing step fails is declined, and the failing step is named.
  • Cross-family verification. In the biophoton pack's first run, both judge families passed their own family's hypotheses more often: the Claude judges by 12 points, the Codex judge by 40. A self-verified certification rate is not a measurement.
AGPL-3.0-or-laterOpen researchModel Context Protocol

Status

3 engines are built and pass conformance on a fictional toy corpus, and 2 of them have run on the biophoton pack. 6 more are designed and wait on an API key or an adapter.

See engines and judges