Measurement method

The living lexicon is a system that is being designed for the word game Sanamahti. It is meant to update the lexicon of the game once a week and to measure how many players know each word. This page sets out the design: what is collected from the game, how candidate words are filtered, what evidence decisions rest on and how familiarity is measured.

Status on 22 October 2026. This page describes a design. The first edition is in use: it is based on the Kotus word list of contemporary Finnish, and the group game and the daily challenge take the words of a board from it. The weekly run, the screen, the cases and the familiarity measurement are not in use yet, and nothing described below is in use in the game unless stated otherwise. The word jury (Sanaraati) is a rating page that can be reached only through an invitation link. The first experiments have been run. The figures are design values and are being calibrated in experiments before launch.

What a swipe is

A player forms a word by swiping across adjacent letters on the grid. If the string is in the board’s word list, the player scores. Otherwise the swipe is rejected. This is how the game works today.

The browser version of the game also runs an earlier trial called Sanapulssi. It counts how many times each rejected string has been swiped. The counter holds no information about the player, the board or the game mode. The counts are indicative and not fit for research. The living lexicon will replace the trial, and the strings the trial has collected are meant to be moved into the new system as its starting data.

In the design, a rejected string is stored as a surface form. The record holds the string itself, the number of swipes and an estimate of how many different players have swiped it. The same string is counted only once per player per round. At the end of a round the player sees their rejected swipes with explanations and can propose any of them as a word with one tap.

A swipe is a hint. It is not a vote for a word. A player can try strings at random, and a round has a time limit. For that reason no figure produced by players can change the status of a word on its own. Swipes and proposals can open a case, but a decision needs evidence from outside the game.

The grid also limits what can be swiped. In a simulation with random boards, a given six-letter word could be formed on about one 4×4 board in 850, and an eight-letter word on about one in 21,000. Swipes are therefore not a sample of the lexicon, and new words are hardly ever found on the grid by chance. That is why new words are meant to come from player proposals and from external evidence.

A morphological sieve before any language model

In the design, every new surface form is first run through rule-based morphological analysers. There are two, Voikko and Omorfi. The sieve places the surface form in one of five classes:

• a headword that is already in the lexicon
• an inflected form whose lemma is known
• a derivative
• a compound whose parts are known
• unknown.

Only unknown forms, and words that players have proposed or challenged, continue past the sieve. The language model therefore sees only the residue.

The order is deliberate. According to the background research, a rule-based analyser identifies Finnish inflected forms reliably, whereas language models let incorrect forms through in measurements. The analyser also gives the same result every time. The assumption will be tested in an experiment before launch.

In the design, the headword decides. A form that has become established as a headword is accepted in the game: for example, the plural-only noun sakset (“scissors”) is a headword in the Kotus word list. The same goes for adverbs such as kotona (“at home”): on 12 October 2026 it was decided that the adverbs, adpositions, pronouns, conjunctions and numerals that are headwords in the Kotus word list will be playable from the first edition of the living lexicon. Other inflected forms are rejected, and at the end of the round the sieve result tells the player which word the form is inflected from. Interjections and a small number of other headwords, such as those with no word class in the Kotus list and some music terms, will be in the dictionary but will not be accepted in the game for now.

Cases and evidence

In the design, a case is opened for an unknown surface form when it has at least four letters and at least five different players have swiped it, or when someone has proposed it as a word. A case is also opened for a word that changes in an update of the Kotus word list, for a decision that a player challenges and for a word on trial when its trial period ends.

Evidence for each case is retrieved from four sources:

• The Kotus word list (Nykysuomen sanalista): whether the word is on the list, and its word class and inflection type.
• Yle’s news archive in the Language Bank of Finland (Kielipankki): occurrences of the lemma per year and a few text snippets.
• Wikimedia: the creation date and citations of the entry in Finnish Wiktionary, and hits in Wikipedia and Wikidata.
• Web search: source, date and snippet.

Three rules govern the use of evidence:

1. Only dated evidence that is older than the case is decisive. Text written after the case was opened cannot settle it.
2. Sanamahti’s own site, AI-generated dictionaries and sources that players can edit do not count as independent evidence.
3. Web search is always supporting evidence. It is never sufficient on its own.

For each piece of evidence the system stores the source, the retrieval time, the source’s own date, the number of hits and a short snippet.

The terms of use of the Language Bank material will be checked before launch. According to the Language Bank, the sentence-scrambled versions of the news archive for 2011–2024 are published under CC BY 4.0, while the versions in the original sentence order are licensed for academic, non-commercial use. The design uses the scrambled versions through the Korp interface. What is still open is whether the terms of the interface allow automated use, and how the source is to be attributed.

The decision

In the design, the language model receives the evidence of the case, the sieve result, the rulebook and the familiarity figures of the word. It returns a structured decision: add on trial, demote, reject, modify or abstain. The decision comes with a rationale and references to the rulebook.

Each decision is checked in three ways:

• A separate checker tries to overturn the decision. If the checker disagrees, the outcome is abstention.
• The rulebook includes tests that a machine can run. They are run on every decision, and a decision that fails a test becomes an abstention.
• Decoy candidates are mixed into every run: pseudo-words, inflected forms and names. If even one of them is accepted, the run proposes nothing for publication.

The rulebook is a text of at most one page. It is derived from the ratings of the word jury (Sanaraati): invited raters answer, card by card, whether a word is acceptable in the game. The AI does not change its own rules or thresholds.

Five rules are fixed and cannot be overturned by the jury’s ratings:

• only headwords are accepted in the game
• proper nouns are not accepted
• words that demean groups of people are not accepted in the game
• web search is never sufficient on its own
• a word can be demoted only on evidence.

Colloquial words, slang, abbreviations and swear words, on the other hand, can be accepted.

Every decision records the model identifier, the prompt version, the rulebook version and hashes of the evidence, so that the decision can be traced. The language model calls will be made through the DeepSeek API. The model has not been chosen yet.

Before the system is allowed to change the lexicon, it is run at least four times in shadow mode: decisions are recorded, but nothing changes. A time-machine test is also planned. It has not been run yet and is not yet one of the conditions for the right to publish. The system would be given an older word list and only evidence that predates a chosen day, and the result would show whether it identifies the words that Kotus has since added to its dictionary.

Alex Liao does not make decisions on individual words. He approves the rulebook, receives a report on every run and can stop publication. The system returns to shadow mode by itself if three reports are unacknowledged, a run crashes, the model changes, a decoy candidate gets through or the cost cap of a run is exceeded.

The weekly edition

In the design, the lexicon changes only once a week. The decisions of the weekly run are first published as proposals, and players can challenge them for seven days. A challenged decision moves to the next run. The others take effect the following Monday as a numbered edition.

For each edition the system stores the changes (words added, demoted, modified and rejected), the number of playable words and a hash of the word set. Each game result records the edition number, so a result can later be tied to the lexicon it was played with.

Each word has a lifecycle status: candidate, on trial, established, rare or rejected. Three further facts are recorded independently of each other: whether the word is in the dictionary, whether it is accepted in the game and whether it is featured on boards. Words are not deleted. A word can be demoted to rare, but its entry, history and rationale are kept.

The familiarity strip

Familiarity is to be measured separately from swipes. A player gets a strip of ten words and says for each one whether they know it or not. Three of the ten are pseudo-words: strings that look like words but are not. How the pseudo-words are generated has not been fixed yet.

A player who claims to know a pseudo-word probably also says yes too readily to real words. For that reason responses are weighted by how the player has responded to pseudo-words, and a guess-corrected familiarity is calculated for the word.

For each word the system stores the number of “know” and “don’t know” responses, the number of pseudo-word responses and the corrected share. The correction formula has not been fixed yet. It will be documented on this page before any figure is published.

The weekly run chooses for the strip the words whose status the automated editor is least certain about. The words on the strip are therefore not a random sample of the lexicon. Answering earns in-game coins. Demoting a word requires at least 200 ratings of that word.

The closest counterpart in the research literature is word prevalence: the share of respondents who recognise a word as a word when real words are mixed with pseudo-words.

Known biases

• The players are not a sample of Finnish speakers. The composition of the player base has not been measured. The players are self-selected, and because the daily challenge has leaderboards for universities and student associations, students are probably more common among them than in the population. No background information, such as age, place of residence or native language, is attached to the observations, so the results cannot be weighted to match the population.
• “I know this word” is a self-report. The response says that the player recognises a written word without context. It does not say whether they can explain the word or whether they use it.
• Strip words are chosen by uncertainty. Coverage is uneven: some words collect many ratings and others none.
• The reward may tempt players to answer quickly. The pseudo-words are meant to reveal careless answering, but three pseudo-words on one strip is a small control. A more accurate estimate of a player’s reliability is available only after several strips.
• The board steers swipes. A player can swipe only what the letters on the board allow, and long words are rare on boards.
• The rules of the game steer attempts. The game accepts only base forms. A player who knows the rule probably tries inflected forms less often, so the number of attempts also depends on the player’s experience.
• The rules limit the data. Proper nouns and words that demean groups of people are not accepted in the game. A headword contains only letters, so multi-word expressions and hyphenated words are left outside the lexicon.
• Meanings are written by AI. The meaning and the example of a word are written by AI and labelled as such. They are not dictionary definitions. Definitions from Kielitoimiston sanakirja, the Kotus dictionary of contemporary Finnish, are not copied or paraphrased.

What is not collected

• No name, email address or other direct identifier is attached to the observations. A player’s trace is stored under a pseudonymous key that is computed on the server.
• The AI does not see the player’s key.
• When a player deletes their account, their trace is anonymised automatically. Aggregate figures and decisions are kept.
• Free text written by a player is not passed to the deciding model as it is.
• Example sentences are not taken from messages of identifiable people.
• The published data contains no player identifiers. The name of the player who first found a word is shown only with that player’s consent.

This site does not use cookies. Its privacy notice is in English. The privacy notice of the game is on the privacy page (in Finnish).

What is still open

• the language model
• how the time-machine test is run and its pass mark
• the guessing correction formula and how pseudo-words are generated
• the terms of use of the Language Bank material for automated use
• the licence of the open data and the content of the first release.

If you see an error in the method, or a better way, say so through the contact form. The data plan is on the data page.