Agentic Q2C: structural recovery and calculation selection in a synthetic nuclear inverse problem

Abstract

Nuclear interaction calibration must distinguish inadequate functional form from poorly fitted couplings. We test a resident language-model investigator that revises the isovector interaction of a spherical relativistic Hartree model. Continuous optimization fits synthetic binding energies, charge radii, and neutron skins; the investigator selects registered density-dependent and mixed-vector terms. Generator couplings and withheld observations are excluded from investigator request packets. In the larger-basis experiment, adding linear and quadratic rho-coupling density dependence reduces the training objective from 27107.26 to 8.18e-21. The recovered couplings reproduce withheld zirconium-90 observables with objective 8.48e-20, and the scaled response Jacobian has full local rank. Substantial resolution shifts nevertheless preclude a claim of converged physical predictions. In a separate calculation-selection pilot with two misspecified interaction families, full- and masked-evidence investigators select the same numerically resolved tin-132 neutron-skin calculation, but both predict the wrong ordering. After acquisition, the linear rival's audit risk rises by 0.54%, while the mixed rival's falls by 75.7% and remains larger. A separate ten-case retrospective evidence-consumption demonstration matches all twenty frozen requested-use decisions, with one explanation qualified by adversarial review; these known fixtures do not establish review efficacy. The results separate useful selection, correct explanation, predictive improvement, and justified reuse. They demonstrate restricted same-solver structural recovery, not global uniqueness, validation against measured nuclei, or an advantage over numerical search.

Keywords