The models are the part you can see. The tools that produce them are the other half of the work, and most of our method lives there. We’d rather earn a capability with better data than hope it shows up in a bigger run.
Why we distill
Scaling gets you more of everything, including the waste. Distillation lets us keep the reasoning and drop the rest. We want concise chain of thought, meaning models that reach an answer by a shorter path, and that only happens if the training data is itself short, correct and dense. Three tools produce that data.
Hadron
Hadron is our distillation framework, built around a tournament that a single teacher runs against itself. The teacher answers, critiques its answer, revises it adversarially and synthesizes the survivors, then ranks the field in a blind Borda vote. The loop repeats until “do nothing” wins twice. The labels that come out beat the teacher’s own single-shot answers, and every role leaves a full reasoning trace that can go straight into process-supervised fine-tuning.
Tokamak
Hadron produces the reasoning and Tokamak makes it lean. It pulls traces out of long, branching conversations and compresses them into a compact stream of internal chain of thought and short outputs. You get dense training data that keeps the reasoning without the sprawl around it.
Stellarator
A distillation run is only as good as the loop around it. Stellarator is the control plane for our fine-tuning and reinforcement-learning runs. A run starts in a sandbox and has to pass a pre-flight check before it scales up, with cost projections, budgets and live alerts at every step.
The data it makes
All of this ends up as datasets. Quark is the first, built for concise chain of thought and token efficiency. It packs more reasoning steps into the same output length. Photon, a family for hardening models around compliance, is in preparation. Repeatable data is what turns one good model into a model line.
