How we distill.

The models are the visible half of the work. The apparatus that produces them is the other half, and it is where most of our methodology actually lives. We would rather earn a capability through better data than assume it will fall out of a bigger run.

Why we distill

Scaling buys you more of everything, including the waste. Distillation lets us keep the reasoning and drop the rest. The goal is concise chain-of-thought — models that reach the answer along a shorter path — and that only happens if the training data itself is short, correct, and dense with signal. Three tools produce that data.

Hadron — tournament refinement

Hadron is our distillation framework, built around a tournament that a single teacher runs against itself. It answers, critiques the answer, adversarially revises it, synthesizes the survivors, then blind-ranks the field by Borda count — looping until "do nothing" wins twice. The labels that come out measurably beat the teacher's own single-shot output. Every role leaves a full reasoning trace, ready to drop straight into process-supervision fine-tuning.

Tokamak — trace compression

Hadron produces reasoning; Tokamak makes it lean. It extracts traces from long, branching conversations and compresses them into a tight stream of internal chain-of-thought and concise output — high-signal training data that keeps the reasoning that got you there without the sprawl that surrounded it.

Stellarator — the control plane

Distillation runs are only as good as the loop you run them in. Stellarator is the control plane for our fine-tuning and reinforcement-learning workloads: sandbox runs pass a structured pre-flight gate before they promote to scale, with cost projections, budgets, and live alerts threaded through every step.

The data it makes

The output of all of this is datasets. Quark is the first — built for concise chain-of-thought and token efficiency, packing more reasoning steps into the same output footprint. Photon, a compliance-hardening family, is in preparation. The apparatus is the point: repeatable data is what turns a good model into a model line.