History and philosophy of science
Inference from singular events
Some events leave many independent traces even though they cannot be repeated. What kinds of inference can be made from those traces?
Research / metascience
Research notes
Can a researcher working with an archive cut off in 1900 derive a discovery made later? The proposed study would test this using historically available methods. The difficulty is keeping later knowledge out of the experiment.
Historical-discovery benchmark / proposal
Testing scientific discovery with a 1900 archive
The investigator would receive scientific publications available at the cutoff date. The test must also account for knowledge already embedded in a model's training and for laboratory equipment that did not yet exist.
Deriving a mathematical result is different from naming a later discovery. A convincing trial would preregister targets, document failed attempts and check whether each proposed experiment could have been performed with available equipment. No such trial has been conducted here.
Archive cutoff / available apparatus / derivations / checks
Research dossier 001 / AI and mathematical discovery / October 2026
Who discovered the theorem?
From program search to formal proofs of open problems, machine contributions to mathematics do not fit a single category. Five source-linked cases follow the different roles played by researchers, models, proof checkers and mathematical judgment.
The dossier includes an interactive contribution ledger: alter what you count as discovery and examine how the evidence changes. The selected cases do not measure AI's share of all mathematics.
Read the illustrated investigation ↗Other proposalsUNFINISHED STUDIES
Metascience
Search and conceptual change
A researcher may need to invent a new representation before any useful hypothesis can be tested. How could a study detect that change?
Forecasting and institutions
Instruments and access
When an experiment requires expensive equipment, it matters who had access to the equipment and who could publish the results.
Philosophy of evidence
Taxonomies and detectors
Two laboratories build detectors for different properties of the same specimens. Their classifications may disagree. To decide which is useful, compare what each predicts about previously untested specimens.
Computational modeling
System identification in an artificial world
Construct a simple world with hidden physical rules. Give investigators instruments and a restricted set of interventions, then compare their inferences with the rules used to build it.
The future of academia
Training in machine-intensive sciences
If software produces most new results, what must students still learn to judge an argument or identify a faulty experiment?