Overtopping failure of non-cohesive earthen dams was investigated in 13 large-scale experiments with dams built of compacted, damp, fine-grained sand. Breaching was initiated by cutting a notch across the dam crest and allowing water escaping from a finite upstream reservoir to form its own channel. The channel developed a stepped profile, and upstream migration of the steps, which coalesced into a headcut, led to the establishment of hydraulic control (critical flow) at the channel head, or breach crest, an arcuate erosional feature that functions hydraulically as a weir. Novel photogrammetric methods, along with underwater videography, revealed that the retreating headcut maintained a slope near the angle of friction of the sand, while the cross section at the breach crest maintained a geometrically similar shape through time. That cross-sectional shape was nearly unaffected by slope failures, contrary to the assumption in many models of dam breaching. Flood hydrographs were quite reproducible--for sets of dams ranging in height from 0.55 m to 0.98 m--when the time datum was chosen as the time that the migrating headcut intersected the breach crest. Peak discharge increased almost linearly as a function of initial dam height. Early-time variability between flood hydrographs for nominally identical dams is probably a reflection of subtle experiment-to-experiment differences in groundwater hydrology and the interaction between surface water and groundwater.