Hypothesis Discovery

Five model entries from the project results supplied on 20 October 2026. Research budgets differ across entries; ranks provide a descriptive ordering by overall Inspire Rate, not a comparison under equal research budgets.

Project results snapshot · Preview. Percentages retain the supplied snapshot's two-decimal rounding.

Ranks are calculated across all models before filtering, in descending order of the reported overall Inspire Rate. Ties share a rank and skip the following position (1, 1, 3). Missing rates and ranks appear as — and are listed last. Filtering and changing the display order preserve these ranks.

Generated counts all hypothesis versions. Filtered counts versions passing literature retrieval. Actionable counts versions judged inspired. Conditional = Actionable / Filtered; Inspire Rate = Actionable / Generated.

By research domain

The domain table uses the overall Inspire Rate rank, not a separate rank within each domain. Domain cells report Inspire Rate in percent. Domain denominators were not supplied; these percentages are not pooled or averaged, and no uncertainty intervals are inferred.

Versions include revisions of an existing hypothesis. Versions in the same research tree are dependent; these denominators are not counts of independent experiments.

What counts as inspired?

A hypothesis version is inspired when a publication after the model's recorded knowledge cutoff supplies a concrete experimental design, comparator, measurement, or testable evidence gap relevant to developing or testing that version.

The connection is correspondence with later research. It does not show that a publication's authors saw or adopted the hypothesis, or that the hypothesis has been experimentally verified.

Every recorded version, including each revision, belongs in the denominator. A version without a positive judgment may lack a qualifying connection or may not have been retrieved. The manuscript therefore treats Inspire Rate as a lower bound on correspondence under its retrieval and judgment procedure.

From a research tree to Inspire Rate

1. 01

   Research tree

   Record every hypothesis version and the version it revises, together with its investigation plan and supporting evidence.

2. 02

   Freeze

   Freeze the research tree and evidence before preparing evaluation materials.

3. 03

   Literature retrieval

   Search the same model-blind reference corpus for each model, restricting references to publications after its recorded cutoff. The best-matching reference is eligible only if it shares at least three non-stopword terms with the hypothesis.

4. 04

   Judge

   An automated judge checks whether the eligible reference supplies a qualifying scientific connection. Identical hypothesis–reference inputs share a judgment.

5. 05

   Inspire Rate

   Divide positively judged versions by all generated versions, including versions without a retrieved eligible reference.

Cutoffs follow the manuscript's recorded operational boundaries: exact days where a serving date is recorded, and month-level boundaries otherwise. Individual cutoff dates and full run configurations are not reproduced in this preview.

The draft describes 6 research databases and 178 expert-proposed starting questions. Agents may develop the supplied questions or set them aside to pursue other directions.

• HCC clinical cohort
• MIMIC-IV
• eICU
• UK Biobank
• GEO
• ArrayExpress

The stated domains are clinical and population health, therapeutic target prioritization, disease mechanisms, and population multi-omics. These are manuscript collection counts; this page does not distribute the research databases or starting-question set.

Results source: two project result screenshots supplied on 20 October 2026. Metric definition and procedure: manuscript evaluation section and Appendix B. This page displays the supplied results; it does not execute a new evaluation.

Benchmark methodology ↗