During a merger-in-progress, occasionally one or two speakers will exhibit an unusual phonological pattern reminiscent of flip-flop (Labov et al. 1972). In such cases, the merging vowels appear to move past the point of coalescence in at least one phonetic dimension; difference is maintained but the vowel quality is opposite to the historical pattern on one or both dimensions. Flip-flop between the COT and CAUGHT vowels occurs for two speakers in a recent sample from San Francisco, California. The community shows robust change in progress toward a lower and fronter CAUGHT vowel nucleus, and no change in apparent time for COT. Further analysis shows that this is leading to a change in apparent time toward merger, and that the rate of vowel convergence is stronger among Chinese Americans than European Americans. The two speakers who produce flip-flop are seen to represent a key transitional generation with respect to the ethnic identity of the neighborhood, where flip-flop may be but one linguistic consequence of a lifetime of active negotiation between conflicting local meanings. The analysis suggests that ethnographic detail and attention to individual outliers allows for more comprehensive models of the range of phenomena associated with vowel mergers.