OpenType. Train a better model, win the duel.
Fine-tune DiffusionGemma, publish the weights, and challenge the champion. A B300 runs both models on the same private cases. Win by a certified margin and the ledger pays your crown.

Train, publish, duel. The ledger does the rest.
Is the model usable in production, not just on one quiz?
Decisions
Typed decisions on short records: a rule, a record, a question. Some record families are sealed and invented per window.
Exact gold from the rule itself; exact odds when the record leaves facts out.
Long context
The same reads buried in long dossiers, 8k to 100k tokens, with corrections and look-alike records.
Exact gold, computed from the dossier, never written by hand.
Ops agent
A multi-turn customer-operations agent with tools and a written policy.
The exact final state the policy demands, replayed by the container.
SQL analyst
A multi-turn analyst that queries a sandboxed database to answer a question.
The exact answer, checked by the container, not by the model.
Paint
A painter that sees its canvas after every turn. “Draw a cat” is a real brief.
Exact pixel checks, or a vision judge scoring the container's own render on a rubric.
This sample of the service measures the decisions track on its levels. The other tracks are listed so you know what to train for; they count once the service publishes them in its plan.
The champion, the ladder, the queue and the crowns.
Each case is scored against its exact gold: half-Brier on a decision, the task loss on an agent or a drawing. Sum the losses on each side and take the log of the ratio. Positive g means the challenger lost less. Across tracks, g is a weighted mix.
A crown needs the lower 99% bound of g, on both halves of the duel, to reach −ln 0.95: at least 5% less loss, with room for chance. No track may regress significantly, and mastered levels must not get worse.
Private cases, exact gold, public audit.
Private, rotating content
Cases come from a secret per window. The teacher's texts for a window stay private until it closes and are never reused.
Sealed families
Part of the decisions are drawn from record families invented for that window alone. Memorising the public ones is not enough.
Gold is never the teacher's
Gold comes from code: rules, policy, SQL, pixel checks. A teacher text is kept only if two other models read back exactly the facts it was written from.
Transcripts are replayed
The container recomputes every agent outcome from the raw model outputs. A worker cannot report a better result than the replay gives.
A side-blind judge
The vision judge sees one render at a time, never whose it is, with a blank-canvas control per rubric.
Commit, reveal, audit
Each window publishes a commitment when it opens and the secret when it closes. A public drand beacon is mixed into the seed. Anyone can re-derive the cases.
Beat the champion. Get the crown on record.
The submit guide has every command: the generators, the publish rules, the signed submit, and how to audit a window once it is revealed.