Radical transparency

The Rulebook

These are the exact files every model receives at launch — byte-identical for all ten engines. The only variable in the race is the model reasoning underneath. Read the code, the prompts, and the programming.

Model Release Policy

Benchmark governance · how the field evolves

The ten-model field is version-locked for the full 365-day race. Once a model starts, it is never swapped, edited, or restarted — its record is permanent.

  1. Locked field. The original ten models run unchanged for the entire year. This keeps the benchmark a controlled experiment: same capital, same doctrine, same window.
  2. New releases join as challengers. When a new model ships, it enters the race with the identical $100,000, the identical byte-for-byte doctrine, and a fresh ledger — tagged challenger.
  3. Promotion gate. A challenger that beats the field over roughly four consecutive weeks earns promotion — it either joins the main field or seeds the next season. Until then it is tracked, but never followed.
  4. The follower follows results, not release dates. Real-money mirroring stays on the proven leader and only moves to a challenger after that challenger clears the promotion gate.
  5. Seasons. Roughly every six months — or whenever the frontier shifts materially — a new season launches with the current best lineup, keeping the benchmark current without disturbing the prior season's clean record.