I begin experiments with Kjartansson's data set with the goal of constructing a near-but-below-surface velocity model. I test lateral extrapolation of data before tomography and propose an alternate definition of pilot trace. Traditionally, a pilot trace is defined as a stack trace. Here I define a pilot trace as the output of a three-dimensional prediction filter. This pilot trace is designed to be insensitive to velocity, dip, and shot statics. The crosscorrelation of this pilot trace with the observed trace is highly sensitive to geophone statics as well as deeper lateral velocity variations beneath the geophones. This crosscorrelation function of midpoint and offset is slant-stacked transforming offset to depth. With suitable normalization the result is a volume of coherency with axes of depth, midpoint, and slowness. The strange ``footprint'' of the 3-D prediction filter confuses the construction of a specific slowness model. A less confusing alternate is a Snell wave approach.