OceanSparx develops AegisMind, a multi-model AI research engine, and publishes its hypotheses, and the results of testing them, at solver.press, including the ones that fail.
An AI research company focused on testable science
OceanSparx's active work is AI-driven research. We build AegisMind, which generates cross-domain scientific hypotheses, subjects them to multi-model debate, and tests those that a computation can settle. We publish the results openly at solver.press.
AI can produce plausible-sounding science faster than anyone can check it. We treat testing as the product: hypotheses are challenged by independent models, and those that a computation can settle are tested against pre-registered protocols. Results are published whichever way they come out, so a refutation is as visible as a success. None of our research leads has yet been confirmed in a wet lab.
Each published hypothesis is labelled as an untested research lead until a test settles it, and comes with an Experimental Validation Package covering methodology, cost estimates and dependency maps, so research teams can see what it would take to test the claim themselves.
A multi-model AI research platform
Five frontier models (Claude, GPT, Gemini, Grok and Mistral) each generate hypotheses independently.
The models critique each hypothesis in structured debate, exposing weak assumptions. This is a filter for weak ideas, not a test of whether a claim is true.
Where a claim can be stated formally it is checked with the Z3 theorem prover, and where a computation can settle it we run it against a pre-registered protocol. Most claims cannot yet be settled this way.
Hypotheses that pass the publication gate are published at solver.press as research leads with their test status; tested results, supported or refuted, are published alongside.
Publication and test record
solver.press is the public window into AegisMind. It publishes cross-domain scientific hypotheses generated and debated by the AI ensemble, labelled as untested until a test settles them, together with the results of those tests and of pre-registered replications of other groups' published work.
Research teams can access full Experimental Validation Packages, including methodology, cost estimates and dependency maps, for any published hypothesis.
Interested in our research, or in collaborating on AegisMind and solver.press research? Send us a message.