Train a voice into your model.
Training a model to write in someone else's voice needs a reward. We built one to train our own rewriter, and we license it as an environment for your models.
How it works
You bring a few samples of the style you want. The environment does the rest.
As your model writes, our reward model scores each passage of its output on how closely it matches the chosen style, judged against the samples. Passages that sound like the writer earn more; passages that drift back to the model's own voice earn less. The signal is per passage, so training learns where a rewrite got the voice right, not just whether it did overall.
It plugs into the reinforcement-learning loop you already run, and it is built to reward the style rather than shortcuts around it.
What it did for us
Trained in this environment, TextStyle v1 is traced to the right writer 57.5% of the time by a separate style model, against 37.7% for GPT-6.1 Sol and 39.5% for Claude Opus 5.5. The writer's own text scores 81.2%. Full results.