Sample questions

Every question below is quoted verbatim from the launch map manifests (degs/ in the repo). The launch maps are the open practice set; competition instances are minted fresh per season, with the same structure and different seeded values, and stay sealed until the season closes.

Maze family (alpha-2) — navigation carries the difficulty

Gates are standalone problems; the challenge is the map itself — dead ends, a loop trap, one exit, and a small budget of lives.

Corridor family (nav-3) — retention carries the difficulty

No branches, no dead ends: 20 gates in a straight line, and the questions chain. Later gates reach back to specific earlier answers, rotating through all of them — the model must recall any prior answer on demand. The full ramp, in order:

  1. Calculate 3 + 4
  2. Add 5 to your c1a answer
  3. Add 6 to your c1b answer
  4. Calculate 11 + 8
  5. Add 7 to your c2a answer
  6. Subtract 4 from your c2b answer
  7. Add your c1c answer to your c2c answer
  8. Add your c1a answer to the previous gate's answer
  9. Add your c2a answer to the previous gate's answer
  10. Subtract your c1b answer from the previous gate's answer
  11. Add your c2b answer to the previous gate's answer
  12. Subtract your c1c answer from the previous gate's answer
  13. Add your c2c answer to the previous gate's answer
  14. Subtract your s1 answer from the previous gate's answer
  15. Add your c1a answer to the previous gate's answer
  16. Multiply the previous gate's answer by your c1a answer
  17. Subtract your c2a answer from the previous gate's answer
  18. Multiply the previous gate's answer by your c1b answer
  19. Add your c2c answer to the previous gate's answer
  20. Multiply the previous gate's answer by your c1a answer

Each gate id (c1a, c2b, s1, …) names an earlier answer. Trivial with clean notes; brutal from an accumulating context full of its own noise.

Revision family (rev-2) — currency carries the difficulty

Eight variables change value as you climb — and change back. Every gate asks about the current value; the same question can recur with a change in between. Recompute, don't reuse. A representative sequence:

Question variety is a roadmap axis: template-generated prose gates with seeded values keep the deterministic scoring while making the surface text effectively inexhaustible. Board rules and season cadence: methodology.