A denominator you can inspect.

What counts as a word?

A lemma is a dictionary-style base form. Inflections such as books and went are mapped to book and go when Princeton WordNet supplies that mapping. Different senses or parts of speech with the same lowercase spelling count once.

Version 1 includes lowercase, alphabetic, single-token lemmas. It excludes proper-name-only entries, numerals, abbreviations without an ordinary lexical sense, multiword phrases, apostrophised forms and hyphenated compounds. Specialist, regional or dated words can remain if they pass those rules and fall inside the frequency cutoff.

Where the 50,000 come from

Surface-form frequencies come from wordfreq 3.1.1, a multi-genre snapshot drawing on sources such as books, subtitles, news, Wikipedia, web text and social media. Lemma eligibility and morphology come from Princeton WordNet 3.0. The resulting snapshot is dated—it is not a live census of English.

The build sums the frequencies of surface forms explicitly mapped to each lemma, sorts those aggregates, and retains the first 50,000. The full compressed dataset, source versions, policy, licence and attribution live together in the project repository. Snapshot: wordfreq 3.1.1 / Princeton WordNet 3.0.

Five measured bands

Twenty questions come from each band. The two rarest bands now represent 15,000 lemmas each, so one answer moves the central estimate by at most 750—not the 2,000 words used by the old 85,000-word prototype.

Scoring

adjusted rate = max(0, (correct − wrong / 3) / 20)
band contribution = adjusted rate × lemmas in band

With four choices, a random guess has an expected score of zero: one chance in four of +1 and three chances in four of −⅓. “I don’t know” contributes zero without a further penalty, and remains in the denominator.

The central estimate directly sums those five band contributions. The conservative estimate sums the 10th-percentile resampled rate in each band, so a lucky isolated success in a rare band is not treated as firm evidence. The plausible sampling range is the 5th–95th percentile of 2,000 deterministic, stratified bootstrap resamples.

What the numbers do not prove

The bootstrap measures sensitivity to this 100-question sample. It does not correct weak distractors, corpus bias, ambiguous senses or editorial mistakes. The question bank has not yet been calibrated on representative participants, so Evokable does not publish population percentiles or claim scientific validity.

Recognition is also only one kind of knowledge. Choosing a definition does not prove that someone can recall, use, pronounce or understand every sense of a lemma.