Darmok is the first automated system to achieve superhuman performance on British cryptic crosswords. Unlike American‑style puzzles, cryptic clues hide an unmarked definition and wordplay behind a misleading surface reading, sparse grids leave nearly half of the letters unchecked, and tournament rules rank any grid with a single wrong or blank square behind all error‑free solutions regardless of speed. In the 55th Times Crossword Championship, Darmok competed alongside elite human solvers, solved all seven championship puzzles without error, and was faster than every human in each round, finishing the final in 6 m 52 s—3 m 13 s ahead of the human champion.
The system uses a three‑model round‑robin cross‑check: each clue is first answered independently by three separate models, then verified crossing letters are used to unlock harder clues. If blanks remain, a whole‑puzzle model together with dictionary look‑ups fills the gaps, but never overwrites confirmed words. This architecture maintained near‑zero errors over a five‑month daily evaluation. Benchmarking shows that removing backup models, fallback steps, or elaborate prompt instructions barely changes overall accuracy, making the design appear over‑engineered. However, on the worst days of operation and during a severe slowdown of the primary model in the live final, the backup paths carried the solve, indicating that once foundation models master the average case, the central challenge shifts to worst‑case resilience.
State‑of‑the‑art models can now solve an entire puzzle in a single attempt with comparable accuracy and speed, but cloud API calls still stall, time out, trigger safety refusals, or get stuck in reasoning loops. By tightening constraints, cross‑validating clues, and layering backups, Darmok achieves tournament‑grade accuracy, speed, and reliability.
Review: Darmok demonstrates that combining redundant models with progressive constraint enforcement yields a robust solution for high‑stakes, low‑tolerance tasks, offering a repeatable blueprint for future complex AI systems.