Editorial Policy

Contexto publishes a puzzle, a set of hints and an answer every day, in several languages. This page describes exactly how those are produced, what is deliberately excluded, and what to do when something is wrong.

How the ranking is built

Each language has its own vocabulary of tens of thousands of words, drawn from a large corpus of ordinary written text and ordered by how often the words actually occur. Every word carries a 300-dimensional vector from a publicly available embedding model, and a guess’s rank is its position when the whole vocabulary is sorted by similarity to the day’s answer.

The ranking is computed once, offline, before a puzzle is ever published. It cannot shift while you are playing, and it is identical for every player.

How the answer is chosen

Candidates are scored automatically on four signals: how common the word is, whether it names a thing rather than serving a grammatical function, whether it is concrete rather than abstract, and how many distinct meanings it carries. Words with one dominant, concrete sense make better puzzles, because the words nearest to them point in one direction instead of several.

The scored list is then reviewed by a person before publication. The automation narrows thousands of candidates to a shortlist; it does not have the final say.

What is excluded

• Slurs and obscenity, via a maintained blocklist, in every language.
• Ordinary words that would still make a poor answer of the day — suicide, weapons, serious illness, and comparable subjects. These remain guessable, so a reasonable guess still receives a real rank; they are only barred from being the answer.
• Proper nouns, misspellings and words from other languages, which are filtered out of the vocabulary itself.
• Plurals and inflected forms as answers. The base form always makes the cleaner puzzle.

Where machines are involved, and where they are not

The vocabularies, the similarity rankings, the hints and the closest-word lists are generated by software. The explanatory writing on this site is written by a person. The answer list is machine-shortlisted and human-approved.

Answers are never chosen to promote anything, and no advertiser has ever influenced, or will influence, which word appears on a given day.

Known limitations

Being straight about this matters more than sounding authoritative:

• The model learns from how people write, so it inherits the associations in that writing — including ones that are arbitrary, dated, or culturally narrow.
• Similarity is not the same as relatedness. Words that are opposites frequently rank close together, because they appear in the same sentences.
• Coverage outside English is thinner. The Portuguese and Spanish vocabularies are built the same way but from smaller resources, so more borderline words survive review.
• Occasionally a rank is simply strange. That is a property of the model, not a bug in the puzzle.

Corrections

If an answer is offensive, ambiguous, plainly wrong, or unguessable, report it to hello@contexto.us.com with the puzzle number. Words confirmed as problems are removed from future rotation. Published puzzles are not rewritten after the fact — the archive is a record of what was played, and silently editing it would make every past result meaningless.

Substantive changes to this policy are reflected in the date at the top.

Independence

Contexto is an independent project, not affiliated with contexto.me or its creators. Where other games are named, it is to explain how this one differs.

More detail on the game itself is on the about page and in the FAQ.