Methodology · Pilot version 1
What your vocabulary estimate means
This short test estimates written receptive vocabulary in word families. It is a learning aid with newly written, uncalibrated questions. Its accuracy has not yet been established with learners.
The reference vocabulary
We use the first twenty 1,000-family bands of Paul Nation's BNC/COCA Level 6 word family lists, version 1.0.0. That is a reference population of 20,000 families. Supplemental lists of proper names, marginal words, compounds, and acronyms are excluded. The lists combine frequency information and pedagogical decisions; they are not a precise ranking of every word in modern English.
A family groups inflections and related derivatives under a base form. A headword's tested meaning is used as a proxy for family recognition. It cannot establish that you know every derivative or every meaning in that family.
How questions are sampled
A reproducible selection chooses four different families from each frequency band, giving a bank of 80. An attempt first uses two per band, for 40 questions. The optional second stage uses the remaining two per band. No additional answers are scored once meanings have been revealed.
The question contexts and choices were newly written for this pilot. They are not the published Vocabulary Size Test. Items have not undergone independent specialist review or learner calibration. Very technical, regional, and dated vocabulary remains in the reference list and can affect results.
Guessing and the size estimate
Each question has four meaning choices plus “I don't know.” A correct answer contributes 1, an incorrect choice contributes −⅓, and “I don't know” contributes 0. This makes uniform random guessing average zero before rounding and limits. Partial knowledge and elimination of choices do not follow this simple guessing model.
Within each 1,000-family band, the average contribution is multiplied by 1,000. We sum the bands, limit the result to 0–20,000, and round to the nearest 100. We do not clip individual bands before summing, which would inflate low scores.
Why the range is broad
A weighted Hoeffding bound gives a conservative 90% sampling range under independent, representative sampling and the guessing model. For 40 questions its margin is approximately 5,200 families; for 80 it is approximately 3,700. Bounds are rounded outward and limited to the reference population.
This mathematical range does not cover biased questions, unfamiliar definitions, related meanings, repeated exposure, or the gap between headword and family knowledge. It is not a validated 90% accuracy guarantee for real learners. Additional questions add evidence; validation is needed before stronger claims or a narrower practical range would be justified.
What the result cannot tell you
- Your CEFR, IELTS, TOEFL, reading age, intelligence, or rank among other people.
- How many words you can spell, pronounce, or use fluently in speech and writing.
- How much vocabulary you know beyond the 20,000-family reference ceiling.
- Whether a small change between repeated attempts represents real learning.
The frequency table reports observed correct answers in broad groups. It does not claim mastery of those groups or predict your comprehension of a particular book.
Privacy and saving results
An essential encrypted browser cookie lets you resume your test. Unsaved attempts are removed after 30 days. Basic first-party funnel records retain an entry category such as Google search or direct visit; we do not store search queries or full referring URLs in these records. Empty visitor records are removed after 90 days.
Saving a result requires signing in to your own account. Saved results remain until deletion, and account deletion removes saved attempts and your practice sample. Quiz and result pages exclude third-party session replay. Sharing is optional and can be revoked; the link exposes only the estimate and its uncertainty, without your identity or answers.
Sources and attribution
Nation, I.S.P. (2017). The BNC/COCA Level 6 word family lists (version 1.0.0). Family data are adapted under Creative Commons Attribution-ShareAlike 4.0. The adapted question data use the same license. Changes include sampling, normalization, and original contexts and choices. No endorsement by the source author or university is implied.
Bank: pilot-1 · Scoring: stratified-pilot-1 · Methodology updated September 10, 2026.