Speaker
Description
Large-language-model applications in heliophysics generally operate outside the software in which analysis is performed: they answer questions about archives, generate scripts, or curate metadata. We report on an agent embedded in a running desktop analysis application, and on the verification practices required to use its output.
SciQLop is an open-source PySide6 environment for in situ space plasma data, embedding a Jupyter kernel and reaching tens of thousands of products from CDAWeb, AMDA, the Cluster Science Archive and SSCWeb through the Speasy library. Its agent layer exposes a single MCP-shaped tool list, shared by several interchangeable backends, acting on the live session rather than on a sandbox: the kernel namespace shared with JupyterLab, the product tree and its path-to-data resolver, plot creation and figure read-back, and detached background jobs for long archive fetches. A second group of tools reaches the published literature, searching arXiv and NASA ADS and retrieving open-access full text from an arXiv identifier, DOI or ADS bibcode. Data and literature are consequently available within the same reasoning step, so a hypothesis can be framed from prior work, tested against archive data, and compared with published results without leaving the environment. Tools that modify state are gated and confirmed per call.
Three studies built with this workflow are presented: a multi-mission survey of solar-wind radial evolution across the heliosphere, in which scaling exponents are measured from spacecraft conjunctions rather than from pooled statistics; an assessment of how well a single L1 monitor represents the solar wind reaching Earth, whose measured decorrelation length independently reproduces published flux-tube widths; and a temperature-anisotropy analysis of Saturn's inner magnetosphere from reconstructed Cassini/CAPS distributions, which returns a negative result at catalogued mirror-mode events.
The transferable outcome is the verification discipline. Results that initially appeared striking were traced to incomplete instrument sampling, to a correlation that survived only until a confounding variable was controlled for, and to a calibration check that was true by construction. We conclude that agreement between an agent's result and expectation carries no information unless the check is independent of the procedure that produced the result, and we describe the practices we now apply by default.