In-Code MAGIC Weight Windows for Deep-Penetration Gap Streaming: A Fully Reproducible Open-Source Benchmark Derived from a 1984 Fusion Shield Study

Deep-penetration quantities in radiation shielding — the dose behind an unbroken shield, or the residual leakage through a well-mitigated streaming path — lie many decades below the source and are the quantities that set design margins. They are also the ones analog Monte Carlo resolves worst. The established remedies, CADIS and FW-CADIS, require an external deterministic transport solve to build the adjoint source, which places them outside a pure Monte Carlo workflow. We show that the mesh-based MAGIC method implemented natively in OpenMC converges such quantities without any deterministic solver, and we quantify the gain on a shielding problem with real design provenance: 14-MeV neutron streaming through straight and stepped gaps between the outboard shield sectors of the STARFIRE conceptual tokamak. For the unbroken-shield background, analog sampling reaches a fractional standard deviation (FSD) of 0.142 at 107 histories while in-code weight windows reach FSD = 0.017 at the same history count — a precision requiring roughly 70 times more analog histories to match; the corresponding factor for the stepped-gap leakage is roughly 50. Weight-window unbiasedness is established by a like-for-like analog comparison, independent-seed replicates, and mesh- and iteration-count sensitivity studies, all agreeing within 0.7σ. The physics of the model is checked against a 1984 MORSE-CG study of the same configuration: the straight-slot transverse dose profile and the three-step staircase result reproduce quantitatively, and the single-step configuration reproduces the expected two-peak shadow-and-redirection structure. The complete model runs end-to-end on a free cloud tier and is archived with a citable DOI, providing an open, reproducible deep-penetration test case for weight-window methods.