Work

Constructing Possibility as an Epistemic Method: An Onturgic Audit of AI for Science

Year: 2026
Type: preprint
Abstract: Does building in AI for science constitute an epistemic method, or merely an instrument forgenerating predictions? This paper develops an onturgic audit—an operationalization of boundedconstructivism and the ONTacture cycle—to answer this question without either promotionalor deflationary bias. The apparatus comprises a possibility graph with agent-relative access,a six-rung deployment ladder paired with a parallel epistemic track (Conjectured–Verified–Understood), a five-rank failure taxonomy with a blind coding protocol, and a six-dimensionalaudit of the possibility horizon. Two conceptual additions are introduced: a threefold typology ofconstruction (object, representation, search process) and the axis of verification severity. Thirteencases and one negative control are audited with explicit trust labels, spanning AlphaFold, Aurora,tokamak plasma control, a syrinx prosthesis, phage vault particles, FunSearch, AlphaEvolve,Co-Scientist, Robin, ART, GNoME, A-Lab, MatterGen, and BEN-2293. Three correction eventsof 2026—the authors’ correction of the A-Lab claim (from 41/58 to 36/57 with tempered novelty),Juelsholt’s peer-reviewed challenge to MatterGen’s experimental validation, and the prioritydispute surrounding ART—reveal a common failure pattern: a training-data echo dressed asdiscovery, rooted in the ontology encoded in the architecture rather than in implementationor parameters. The evidence supports construction as a method of search, but as a method ofknowing only conditionally: construction yields knowledge where verification is independent,cheap, and severe. Three pre-registerable tests with explicit rejection criteria are specified
Language: English
Open Access status: green
OpenAlex ID: W7225888244
Cites:
Cited by:
Related to: 10