Local scientific investigators on distributed compute islands

Abstract

Large scientific parameter spaces require more than efficient evaluation of a fixed model. A stalled search can indicate a poor coordinate system, insufficient numerical accuracy, missing physics, or a software defect. These possibilities require different experiments. We propose an architecture in which a small resident language model directs fast local investigations within a leased compute island, while a stronger coordinator occasionally reviews evidence, changes scientific assignments, and reallocates resources. Numerical refinement and inference remain the responsibility of explicit algorithms. The investigator chooses what those algorithms should investigate and when their assumptions need revision. As an executed proof of concept, a Gemma 4 E2B model on an Open Science Grid GPU confronted a capped black-hole birth-mass model with released GW190521 samples. It made 3 model calls and selected 5 experiments beyond a supplied baseline. The local loop completed in 16.39 seconds within a 300-second successful allocation. Enlarged mass envelopes accommodate much more of the event posterior, but an exact equivalence test rejects the agent's claim to identify a formation channel. This failure makes scientific adjudication part of the architecture. We specify topology-aware allocation, experiment contracts, evidence exchange, and a staged evaluation program, then develop a hydrodynamic transition-mapping example and a longer discussion of scale, model revision, and computational discovery. The demonstrated result is a remote evidence-to-experiment loop; multi-island coordination, hydro investigations, and an efficiency advantage remain to be tested.

Keywords