Synthetic Data That Survives Contact With Reality

Synthetic data generation has an obvious economics story and a much harder quality story. Unfiltered generation reliably produces distribution collapse: the model learns the generator's biases faster than it learns the task.

Our pipeline treats generation as a proposal step and spends most of the compute on filtering. Physical plausibility checks, diversity scoring against an embedding index, and a small human-labeled reference set that acts as a distribution anchor all sit downstream of the generator.

We target a fixed synthetic-to-real ratio per task rather than maximizing synthetic volume. Past roughly 5:1 in our vision tasks, additional synthetic data stops helping and begins to erase rare real-world modes.

The rare modes are the point. Most physical deployments fail on the tail, and the tail is exactly what naive generation smooths away.