TL;DR
Darmok is the first AI system to achieve superhuman performance on British cryptic crosswords, winning the 2026 Times Crossword Championship by finishing over three minutes ahead of the human champion. Its success stems not just from raw LLM reasoning, but from a "resilient orchestration" architecture designed to maintain zero-error performance even when primary cloud APIs suffer from latency "brownouts" or reasoning loops.
What Problem Does This Paper Address?
Crossword solving has long been a benchmark for AI, but British cryptic crosswords represent a significantly higher bar than American-style puzzles. While American clues are largely semantic queries (e.g., "Capital of Peru" → LIMA), cryptic clues are "dual-grammar" puzzles. They contain a misleading "surface reading" that sounds like a normal sentence, but they also function as a "string program" providing specific instructions (anagrams, deletions, reversals) to construct the answer.
Furthermore, British grids are "topologically sparse," meaning nearly half of the letters are "unchecked" (they don't intersect with another word). In a tournament setting, a single wrong letter ranks you behind every perfect solution, regardless of speed. The challenge is building a system that is both lightning-fast and perfectly accurate, while relying on volatile, heavy-tailed cloud reasoning APIs.
Methods
The authors introduce Darmok, a multi-model ensemble that treats crossword solving as an iterative, synchronous constraint-tightening loop. Key technical innovations include:
- Multi-Seat Orchestration: Darmok queries three independent models (Gemini 3.1 Pro as the "Anchor," Claude Fable 5 as the "Peer," and GPT-5 as the "Third Seat") in parallel.
- Synchronous Pass Barriers: Instead of "greedy" admission (writing an answer as soon as it's generated), Darmok uses a "rebuild-from-kept" recurrence. At the end of each pass, it clears the grid and only admits answers that are corroborated by multiple models or strongly supported by crossing letters.
- Jacobi-Style Updates: The system uses a scoring mechanism where words gain "standing" (tentative, supported, or locked) based on agreement with intersecting neighbors. This prevents unvetted guesses from "bootstrapping" each other into the grid.
- Resilient Failover: A background "Boundary Feed" (GPT-6-astra) attempts to solve the whole puzzle in one shot. If the primary per-clue models stall or time out, these background answers are used as non-evicting backups.
Datasets
The system was evaluated across three primary corpora:
- Post-Cutoff Daily Production Record: 132 daily Times puzzles (3,876 clues) published after the training cutoffs of all constituent models to ensure no memorization.
- Stratified 16-Puzzle Test Set: 463 clues spanning five difficulty bands (Easy to Extreme) on the community "Snitch" scale.
- Mephisto Transfer Set: 10 "barred" puzzles (360 clues) featuring zero black squares and archaic vocabulary from the Chambers Dictionary to test the system's generalization to different grid topologies.
Key Insights and Ideas
- The "Bitter Lesson" of Resilience: As foundation models master the "average case," the central engineering challenge shifts to the "worst case." Under normal conditions, Darmok’s architecture looks over-engineered. However, during the live Championship Final, the primary Gemini API suffered a 72% drop in token delivery speed. Darmok’s backup paths carried the solve, while a "naive" single-model solver would have timed out.
- Crossings as the Primary Reasoning Amplifier: The researchers found a "saturation jump" in accuracy. Single-clue solve rates are mediocre (~50%), but once a clue has 3 or more confirmed crossing letters, solve rates jump to >90%.
- Prompting as a "Brake": Elaborate Chain-of-Thought (CoT) prompting didn't actually improve the models' reasoning accuracy. Instead, requiring a wordplay explanation acted as a "brake" on unverified guessing, converting would-be errors into "clean" timeouts that the system could then handle via failover.
Key Results
- Superhuman Performance: In the 55th Times Crossword Championship, Darmok solved all 7 puzzles (210 clues) error-free. It cleared the Final in 6m 52s, beating the human champion by 3m 13s.
- High Reliability: Across the 43-day "frozen" production epoch, Darmok achieved a 95.3% perfect-grid rate and 99.84% clue accuracy.
- Zero-Shot Generalization: Without any modification, the system achieved 97.2% accuracy on the notoriously difficult Mephisto barred grids, which use a completely different grid geometry.
Limitations and Future Work
- Thematic Meta-Puzzles: The system still struggles with "The Listener" style puzzles, where a prose preamble defines new, puzzle-specific rules (like ciphers or hidden instructions). These require zero-shot synthesis of algorithms rather than just solving clues.
- The "20-Watt Wetware" Gap: While Darmok is superhuman, it requires megawatts of datacenter compute and high-speed internet. A human champion uses ~20 watts of "biological wetware" and a pencil. Developing a fully local, air-capped solver is the next major systems challenge.
- Cryptic Setting: Solving is "recognition"; setting (creating) witty, valid cryptic clues is the "Olympus Mons" challenge for LLMs due to the need for simultaneous surface fluency and structural soundness.
Summary updated