Transparent by design

Wordle helper methodology

The constraint engine treats position rules and letter counts separately, then combines every completed guess before filtering candidates.

Word lists

The allowed-guess dictionary contains 8,636 five-letter entries derived from the public-domain ENABLE word list. The stored source snapshot has SHA-256 61ba1392a5b6199dd161fae7b483cd262fb8a88c0d49bd2c9ecb6753e755697f.

Commonness scores are derived from wordfreq 3.1.1 by Robyn Speer and contributors. The generated frequency data is offered under CC BY-SA 4.0. The interface keeps likely answers and uncommon valid dictionary words visibly separate.

The likely-answer layer is a commonness heuristic, not a copy of a private or guaranteed future answer schedule. Proper names, offensive terms, abbreviations, and unusual inflections still require continued editorial review.

The full candidate review uses Princeton WordNet as a lexical screening aid and the U.S. Social Security Administration's public name data as a name signal. Neither source automatically excludes a word: familiar ordinary senses are retained, and every exclusion has an explicit, reversible reason.

Reveal separates matching words into Common, Possible, and Dictionary groups. Common and Possible come from the likely-answer layer; the split at a 0.60 normalized commonness score only makes the list easier to scan. Dictionary words match the tiles but fall outside that layer. None of these labels is the probability that a word is today's answer.

How the next guess is ranked

Each guess is tested against every remaining candidate using the same repeated-letter scoring rules as the board. The helper measures how evenly that guess divides the candidates, its expected and worst-case number of words left, whether it can be the answer, and whether it is common enough to be useful in practice.

Filtering and ranking run locally in a background Web Worker. This keeps the board responsive while the full dictionary is evaluated and avoids uploading puzzle input for core calculations.

Normal Mode may recommend an information guess that is not a possible answer. Hard Mode filters those recommendations through every known position and letter-count constraint. The result card labels the recommendation type and shows the worst-case remainder instead of presenting an unexplained score.

When several likely answers share a one-position pattern such as _IGHT, the helper labels the word-family trap. Normal Mode rewards guesses that test several of the changing letters together; Hard Mode explains when its clue requirements prevent that strategy.

Full-game validation

Recommendation quality is tested as a complete game, not only as a good-looking first guess. The benchmark starts with RAISE, generates the exact gray, yellow, and green feedback for every word in the reviewed likely-answer layer, and follows the helper's recommendation until the answer is solved or all 6 turns are used.

Answers with the same clue history share one solver calculation. This feedback-state tree covered all 1,401 likely answers while still checking that each real answer survived every filtering step.

Normal Mode solved all 1,401 benchmark answers, with no dropped answers, missing recommendations, or repeated guesses. Hard Mode solved 1,397 of 1,401. Its unresolved cases—HATCH, SHAVE, JOKER, CRORE—sit in families where obeying every revealed clue can force sequential guesses.

Machine-readable reports: opening comparison, feedback-pattern quality, and complete-game simulation.

What the benchmark does not prove

The test measures this documented word snapshot and solver version. It does not prove that every word in the likely-answer layer can appear in the official game, and it does not predict a private daily answer. A word-data or ranking change must rerun the same benchmark before these figures are updated.

Limits

This independent tool does not claim access to an unpublished New York Times answer schedule. A recommendation is a calculation over the documented local word data, not a prediction of a privately selected daily answer.

Past answers are not removed by default. Wordle now permits earlier answers to return, so “already used” is a preference rather than a logical constraint. The solver supports an optional, dated history snapshot, but the production build does not bundle an unreviewed third-party answer archive.