def split_dataset(rows, ratios, group_key, seed):
    return {"train": [...], "validation": [...], "test": [...]}

1. rows are grouped by group_key (the author field)
2. no group appears in more than one split
3. every row is assigned exactly once
4. split sizes follow the ratios approximately
5. the same seed produces the identical split every time
6. ratios must sum to 1

• group membership is decided by hashing the group value with the seed
• a group is never split across boundaries
• the input rows are not mutated