A 243,291-row supervised fine-tuning curriculum designed to produce honest, direct, and agentic language models. The "Robust" designation refers to the intentional diversity of system prompts, reasoning styles, and source provenance — training models that hold up across distributions, not just within one.
