Darmok’s cryptic crossword sweep puts AI backup systems to the test

Darmok, an automated cryptic crossword solver, completed all seven puzzles at the 2026 Times Crossword Championship without an error, according to a new preprint by Siddhartha Chatterjee of the University of Texas at Austin. Its final grid took 6 minutes 52 seconds, finishing 3 minutes 13 seconds ahead of the human champion’s error-free submission.

The system participated with the organizers’ permission, outside the official human prize standings. Its reported result is striking, but the paper also complicates the achievement: one standalone model subsequently solved every championship puzzle correctly across three attempts per puzzle. Darmok’s elaborate architecture was not necessary to match that accuracy in those trials. Its intended advantage was protection against failures that an ordinary benchmark might miss.

Why crossing letters are not enough

British cryptic clues usually contain both a definition and wordplay that constructs the same answer. The wording conceals where one ends and the other begins. A solver might need to reverse an abbreviation, insert it inside another word, or rearrange selected letters while ignoring a misleading surface meaning.

Unlike standard American crossword grids, British cryptic grids leave many letters without a perpendicular crossing. Two possible spellings can therefore fit every crossing while only one satisfies the clue. Completing a consistent grid does not establish that every answer is correct.

Championship scoring makes those remaining uncertainties costly. Any imperfect submission ranks behind every perfect one, regardless of speed. Darmok consequently needs more than a stream of plausible answers: it needs rules for deciding which proposals to trust and how to remove mistakes without damaging other entries.

Accepting whole words, not overwriting squares

Darmok sends individual clues to models from Google, Anthropic and OpenAI. At the end of each solving round, it evaluates their proposals together rather than accepting whichever response arrives first. Agreement between models, model-specific confidence thresholds and support from previously established crossing answers determine which words enter the grid.

The system then rebuilds the grid from the accepted words. Conflicting proposals with equal scores are both withheld. This avoids a failure seen in an earlier version, where replacing individual crossing letters could turn an accepted answer into a non-word while leaving it marked as solved.

Unresolved clues return to the models with tighter letter patterns and, where available, dictionary candidates. A separate background model attempts the entire puzzle. That process and local dictionary-based recovery steps can fill gaps without overwriting accepted answers at the point of insertion.

The final also used three parallel solving replicas, submitting the first to finish. In the earlier three-puzzle rounds, the system solved different puzzles concurrently; the paper says the sum of their individual times was the comparison arrangement agreed with organizers. These are not equivalent computing conditions to a person working through paper grids.

When the primary model slowed

During the final, the paper reports that Darmok’s primary model timed out on 23 of 30 opening clue requests in the winning replica. Client logs and project-level cloud measurements showed a substantial service slowdown compared with earlier rounds.

The backup paths supplied 17 answers without the primary model: 13 came through the other individual-clue models and four from the whole-puzzle process. Together, they completed the grid without an error.

That protection also imposed a delay. The whole-puzzle model had already returned a correct solution at 5 minutes 38 seconds. Darmok waited for its next synchronization point before using it, finishing roughly 74 seconds later with additional cross-checking.

The longer evaluation provides a less dramatic reliability measure. Across 43 daily puzzles using the frozen championship configuration, Darmok answered 1,259 of 1,261 clues correctly, but only 41 grids were flawless. On ten harder Mephisto puzzles, which use obscure vocabulary and barred grids, it produced five perfect grids. Several errors involved spelling variants that crossing letters could not distinguish.

What the evidence leaves open

The work is a preprint, and its tournament account and measurements are author-reported. Several model identities appear under aliases described as vendor-embargoed, and the manuscript says its repository will be made public upon release, limiting independent inspection from the supplied material.

Seven championship puzzles, even with repeated model attempts, cannot establish a general failure rate. Some component-removal experiments replayed recorded responses rather than rerunning the changed system, so they do not capture how different intermediate grids would change later model answers.

The replicas also shared a host, API quotas and a background model, leaving common failure points. Darmok demonstrates a reported success under a real slowdown, not immunity to outages. Its practical contribution is a concrete way to combine strong but fallible models while keeping their shared crossword state consistent.