Darmok sweeps Times Cryptic Championship: 7/7 error-free, Final 3m13s ahead of human champion

Darmok: Championship Cryptic Crossword Solving via Constraint Tightening and Resilient Orchestration

Siddhartha Chatterjee

The University of Texas at Austin

cs.AI, cs.CL, cs.DC

2026-11-10

Darmok, a three-model redundant solver, solved all 7 Times Championship puzzles error-free, beat every human in every round, and finished the Final 3m13s ahead of the champion.

What problem this solves

British cryptic crosswords break both properties that made American crosswords tractable for machines. An American grid checks every white square in both directions and every clue is a direct semantic query, which is why constraint-satisfaction solvers like the Berkeley Crossword Solver reached superhuman tournament play years ago. A cryptic clue reads two ways at once: as ordinary English it is a fluent, deliberately misleading sentence; parsed under the compositional rules of Crossword English it splits into a definition plus an executable wordplay expression (anagrams, containers, reversals, homophones) that must build the same letter string. Grids are also sparse, with 40% to 50% of each word's letters unchecked, so crossing constraints cannot rescue a misparsed clue. No automated system had previously approached championship human play on full grids under a live clock.

Tournament rules add a second axis. One wrong or blank square ranks a grid behind every error-free solution regardless of speed. At 97% per-clue accuracy, the probability of a clean 30-clue grid is 40.1% (0.97 to the 30th); at the 2026 qualifiers the fourth-fastest solver dropped to 61st place on an 89/90. Human champions clear the Final in 8 to 12 minutes, so a model that thinks for 15 minutes loses the speed race even when it is right. Zero errors, top speed, and reliability must hold simultaneously.

Method

Darmok coordinates three per-clue model seats (Gemini 3.1 Pro as anchor, Claude and GPT-5 as checkers) plus an asynchronous background process that sends the entire puzzle to GPT-6 in one prompt. Each seat gets an empirically calibrated confidence floor: Gemini's self-reported confidence is essentially binary, while Claude shows a sharp accuracy knee at 85.

Solving is an iterative constraint-tightening loop. Pass 0 dispatches every clue cold and in parallel; at each pass barrier a synchronous admission rule scores what returned. An anchor proposal or two-seat agreement earns a base score of 2; a lone secondary proposal scores 1 and needs a lift from an already-admitted crossing word. Crossing support follows a Jacobi-style update against neighbors' prior-pass standings, so two unvetted guesses arriving in the same pass cannot vouch for each other, and tied conflicts drop both words. The grid is cleared and rewritten from scratch at every barrier, which kills the arrival-order cascades and partial-word corruption of the 2025 greedy build; in paired tests on 40 hard puzzles the switch gained 18 clues and cut API cost by 41%.

Admitted words project crossing letters, and open clues are re-dispatched with tightened wildcard patterns like P?U?E?I?E. The design bets on a measured regularity: cold single-seat solve rates run 44% to 64%, jump to 91% to 94% with 3 known crossing letters, and reach a 100% three-seat union at 4 or more. Crossing letters are the reasoning amplifier here, not prompt engineering.

Every backup stage is strictly non-evicting: the whole-grid feed only fills open slots and never overwrites a confirmed word, and a local dictionary chain fills whatever remains, since a blank scores exactly as badly as an error.

Results

SettingMetricResult
2026 Times Championship (7 puzzles, 210 clues)Accuracy / speed210/210, fastest in every round; Final 6m52s, 3m13s ahead of champion Hughes (10m05s)
132 daily Times puzzles (3,876 clues)Clue accuracy / perfect grids99.17%, 108 perfect days; frozen 43-day epoch 99.84%, 41/43, median 5m36s
Stratified 16-puzzle test set (463 clues)Clue accuracy99.57% (median 4m11s); 2025 build 82.1%, June prototype 88.1%
Zero-shot transfer to barred Mephisto gridsClue accuracy350/360 (97.2%)
Standalone GPT-6, whole puzzle in one prompt21 draws21/21 perfect, median 4m32s

Post-championship ablations cut the other way. Removing the backup seats, fallback stages, or the elaborate prompt changes benchmark accuracy by less than replicate noise, and a single anchor seat with crossing retries matches the full ensemble at 461/463. But seconds after the live Final began, the Gemini anchor hit a server-side brownout: median request latency jumped from 16s to 50-190s, token throughput fell 72%, and 23 of 30 opening clues timed out. Backups supplied 17 of 30 answers (56.7%), and the grid still locked error-free at 6m52s. Across the worst decile of the 132 daily runs, overall accuracy held at 98.68%, with 90% of committed answers coming from backups on the peak outage day. One fairness note: the three round puzzles were solved concurrently, so round time is the slowest puzzle (4m20s in the qualifiers against a 12m08s sequential human sum).

Why it matters

Two layers. On the domain fact, frontier models now solve championship cryptics in a single shot at parity accuracy and speed: standalone GPT-6 scored 21/21 perfect draws on the same seven puzzles, and during the live Final its background feed wrote a complete 30/30 at 5m38s, 74 seconds before the per-clue pipeline finished verifying. Another hard reasoning benchmark has effectively fallen. The systems argument is the paper's real point: static benchmark ablations systematically misprice redundancy. Cloud reasoning endpoints are heavy-tailed components whose failures (timeouts, safety refusals, token-budget loops, brownouts) are invisible on calm days and decisive on bad ones. The logic matches RAID and error-correcting memory: zero marginal gain when nothing fails, everything when something does. Anyone assembling multi-model systems over cloud APIs can use that lesson.

There is also a handy set of negative results: for models with native extended thinking, elaborate CoT scaffolds, self-revision fields, and domain tutorials produce no detectable accuracy gain, and requiring a wordplay explanation before the answer mostly converts wild guesses into timeouts.

Limitations

The paper is candid: a 7/7 sweep is a small sample, its own Clopper-Pearson bound puts the puzzle-level failure-rate upper bound at 34.8%, and the reliability claim rests on the 43-day daily record (41/43 perfect grids). All seats ran as vendor-embargoed pre-release snapshots, so outside readers cannot reproduce the behavior.

Three things look weaker on close reading. First, part of the tail-risk evidence was self-inflicted: 9 of the 13 worst decile days came from local macOS DNS exhaustion dropping the anchor's calls, a bug on the author's side rather than the provider's, and the only live provider brownout besides the Final never materialized in that set. Second, the three Final replicas shared one host and one API project quota, which the paper itself notes gives no statistical independence against project-wide failures, exactly the scenario redundancy is meant to cover. Third, the entry was an exhibition outside official standings, with single-author, self-reported telemetry.

Terms

Source

What people are saying

Related papers

All paper explainers