The relationship between seismic velocity and internal fault structure was investigated through a shallow seismic refraction experiment across the Punchbowl fault, Devil's Punchbowl Los Angeles County Park, California. The Punchbowl fault is a northwest-striking, large-displacement fault of the San Andreas system that is exhumed to several kilometers depth and places crystalline basement against arkosic sandstone of the Punchbowl Formation. Seismic refraction profiles using hammer and impulsive shear-wave sources along a 300-m-long line reveal the velocity structure of the fault zone beneath a thin deposit of alluvium. We determine a velocity model assuming the alluvial layer is fairly uniform in velocity and thickness consistent with geologic observations and P-wave travel times. Raytracing with damped least-squares inversion of travel times of P and S waves indicate that the Punchbowl fault is best modeled as a zone several tens of meters wide with velocities reduced by 10%–25% from wall-rock velocities (Vp = 3.2 km/sec for granitic basement and Vp = 2.9 km/sec for Punchbowl Formation). Thickness of the low-velocity zone and the variation in seismic velocity across the zone are qualitatively consistent with expectations based on the observed distribution of fault-related fracturing and alteration. Apparent crack densities calculated from measured seismic velocities using O'Connell and Budiansky (1974) formulation for a cracked medium range from about 0.4 in the core to a background crack density of 0.1 in the host rock. The variation in calculated crack density across the fault is similar to observed variations in microfracture density in the Punchbowl Formation sandstone along traverses across the fault. An estimate of the Poisson's ratio near the fault is about 0.25, suggesting that open cracks in the shallow part of the Punchbowl fault zone are dry, consistent with the geologically inferred location of the groundwater table. Although the seismic data do not completely constrain the velocity structure, the seismic velocity model determined by raytracing and inversion of travel times is admissible on the basis of structural data.