Speaker
Description
Extracting light-cone parton distributions from lattice QCD within Large-Momentum Effective Theory (LaMET) requires a multi-stage data-reduction pipeline: fitting Euclidean correlators for hadronic matrix elements, renormalizing the resulting matrix elements, extending and Fourier-transforming them to momentum space, matching to light-cone distributions in perturbation theory, and extrapolating to the continuum, infinite-volume, and physical-mass limits. Several of these stages are procedurally fixed once a scheme is chosen, but others require a physicist to judge fit windows, ansatz choices, and model-averaging strategies from the data itself, which makes the full pipeline slow to execute and difficult to reproduce across practitioners. We present \lamet, an open-source large language model (LLM) agent framework that formalizes this pipeline as a declarative, job-based manifest and exposes each analysis stage as a registry of typed numerical tools. A tool-calling LLM loop drives the stages that require strategy selection, guided by stage-local ``skill'' prompts that encode the underlying physics heuristics, while stages that are procedurally fixed run as deterministic tool pipelines. We describe the resulting architecture, validate the agent-in-the-loop correlator-fitting stage on synthetic two-point-function data, and outline its application to an end-to-end pion quasi-parton-distribution-function (quasi-PDF) workflow in the Coulomb-gauge formulation of LaMET. This paper documents the design and first release, \lamet\ v1.0, of the framework.
