The written
record.

Abstract

We present a fully autonomous pipeline in which candidate gravitational Lagrangians are proposed, symbolically constrained and confronted with cosmological likelihoods without human intervention. Across 1.4 × 10⁶ candidates, 63 survived dimensional, causal and weak-field consistency filters; four exceeded the ΛCDM evidence baseline.

Abstract

Machine-proposed w(z) parameterisations are ranked by Bayesian evidence rather than goodness of fit. We find a family of relaxation models that mildly prefer phantom crossing near z ≈ 0.7, reducing the Hubble tension to 2.6σ without invoking early-universe modifications.

Abstract

Scientific credibility of automated discovery depends on traceability. We describe a derivation-graph representation in which every accepted theory retains its assumptions, intermediate transformations and rejected siblings, enabling full replay of the reasoning that produced it.

Abstract

We formulate theory exploration as a compute market and allocate accelerators by expected reduction in cosmological parameter entropy. The scheduler yields a 4.3× improvement in discovered-evidence per GPU-hour over uniform allocation.

Abstract

Generative models readily produce dimensionally invalid physics. We introduce a tensor-algebraic verification layer enforcing gauge invariance, causality and correct classical limits, rejecting 99.98% of proposals prior to any numerical simulation.