An instrument for
generating theory.

Every candidate is compiled into a symbolic representation and checked for dimensional consistency, gauge invariance, causality and known-limit recovery. Theories that cannot reduce to General Relativity in the weak field are discarded automatically.

Surviving theories enter an adaptive simulation loop. Bayesian optimisation allocates compute toward parameter regions with the highest expected information gain, producing posterior surfaces rather than single fits.

Predicted observables are confronted with survey data: CMB anisotropy spectra, weak-lensing shear maps, baryon acoustic oscillations, pulsar timing residuals and type Ia supernova distance ladders.

Each theory carries a full provenance graph: assumptions, derivations, contradicting evidence and lineage from prior candidates. Researchers can traverse why a model was proposed, not only that it scored well.

The engine runs on a scheduler that treats theory space as a compute market — allocating accelerators to the branches with the steepest expected reduction in cosmological uncertainty.