While modern agriculture relies on genetic homogeneity in the field, some farmers grow genetically heterogeneous crops and exchange seeds. Such diversifying practices associated with seed recycling may allow adaptation of crops to their environment. This socio-genetical model constitutes an original experimental evolution design called On-Farm Dynamic Management (OFDM). Studying OFDM can help understanding how evolutionary mechanisms shape crop diversity submitted to diverse agro-environmental conditions. We studied a farmer-led initiative where a mixture of four French wheat landraces called "Mélange de Touselles" (MDT) was created and distributed within a farmers' network. Fifteen populations derived by farmers from the initial mixture were sampled after 2 to 7 generations of cultivation on their farm. Twenty-one space-time samples of 80 individuals were genotyped using 17 microsatellites markers and characterized for their heading date in a "common-garden" experiment. Gene polymorphism was studied using four markers located in earliness genes. An original network-based approach was developed to depict the particular and complex genetic structure of the landraces composing the mixture. A rapid differentiation of the mixture was detected, larger at the phenotypic and gene levels compared to the neutral genetic level, indicating a potential divergent selection. We identified two interacting selection processes: variation of the mixture component frequencies and evolution of the within-variety diversity, that shaped the standing variability available within the mixture. These results confirm that farmers' practices increase genetic diversity and allow crop evolution, which is critical in the context of global change. OFDM appears as a promising model of crop experimental evolution.