How the assessment works, in detail

In short: the assessment estimates how far down the list of the 10,000 most frequent Spanish words your solid understanding reaches. It asks you to rate words, picks each next word where your answer tells us the most, and shows the result as a range with its uncertainty. It is self-rated and it measures recognition. It is not a grade and not a DELE score.

This page describes exactly what the app does today, with the numbers it uses. When the method changes, this page changes with it.

1. The word list

Words come from doozan/spanish_data, an open dataset built from Wiktionary, with frequencies from film and TV subtitles (OpenSubtitles, via the FrequencyWords project).

• Lemmas, not forms. Forms are grouped under their dictionary word: habló, hablamos and hablando all count as hablar. Rank 1 is the most frequent lemma.
• Spoken register. Subtitles reflect everyday spoken Spanish. Words common in newspapers or academic writing can rank lower than their importance for DELE reading tasks. A written-register list is on the roadmap.
• Why 10,000. Research on English suggests readers need roughly 8,000–9,000 word families to read unsimplified text comfortably (Nation, 2006). The 10,000 most frequent Spanish lemmas are where advanced learners’ gaps live, and where most courses stop teaching.

2. What you are asked

Each word takes one or two yes-or-no questions:

Example sentences come from Tatoeba (CC BY 2.0 FR) and are shown only if they pass our quality checks in both Spanish and English.

3. The model: your knowledge curve

We assume the chance that you know a word falls as words get rarer, along a smooth S-shaped curve. Two numbers describe your curve: where it falls, and how steeply. Some learners know nearly everything up to a point and little after it. Others fade gradually over thousands of words.

• Where: any point from 0 to 18,000, in steps of 100 (a slowly fading curve can reach past the end of the list).
• How steeply: seven widths, from a sharp drop over a few hundred words to a slow fade over nearly ten thousand.
• Room for mistakes: we assume a 5% chance that you answer “don’t know” to a word you know, and an 8% chance that you answer “know” to a word you don’t (a lucky guess, or a word that looks like English).
• Shaky counts as don’t know in the estimate: a word that was clear only with the sentence isn’t one you know solidly. Your gap list still marks it shaky, separately.
• Each answer counts for less than a full observation (a weight of 0.4). Self-ratings are noisy, so no single answer can swing the result.
• Starting point: before your first answer, we assume a very wide spread centered on 4,000 words. It is wide enough that a handful of answers outweighs it.

After every answer, the app updates the probability of each of the 1,267 possible curves (181 positions × 7 steepnesses) with Bayes’ rule. This is exact arithmetic, not a rule of thumb, and it runs in your browser.

4. Your number and your range

• Your number is the word rank where your chance of knowing a word drops to 80%. Well before that point you know nearly every word; around it, about 8 in 10; beyond it, fewer and fewer.
• Your range runs from the 10th to the 90th percentile of the estimate. Given your answers, the true point lies inside it about 80% of the time.
• The single number (on the share card, for example) is the middle of that estimate, the median.
• Your pattern puts it in words: solid up to your number, patchy until the first block of 1,000 words where the model puts your chance of knowing a word below 50%, mostly new after that. The colored strip shows the same three zones along the list. The block chart shows the model’s estimated share of known words in each block of 1,000: green at 80% or more, yellow from 50% to 80%, red under 50%.
• Numbers are rounded to the nearest 100. If the top of the range reaches 10,000, we say “to all 10,000”: the list ends there, so the test can’t see further.

5. Choosing the next word

• The app considers 100 candidate points along the list (ranks 50, 150, 250 … 9,950) and picks the one where your answer is expected to tell us the most about your 80% point (the largest expected information gain about it).
• It never asks more than 6 words within 250 ranks of each other, so the test doesn’t drill one spot.
• At the chosen point, it picks a word at random within 50 ranks (widening to 150, then 500, if none fits) that has a vetted example sentence. Words you have already rated come last.

6. When it stops

• You need at least 8 answers for a result.
• The app suggests stopping when the range is 4,000 words wide or narrower, or after 25 answers, whichever comes first; and again at 3,000 or after 35. If your knowledge fades slowly, the range can stay wider than 4,000 even at 40 answers. That is an honest picture of a gradual fade, not a failure of the test.
• It stops at 40 answers in any case. Up to 10 skips are allowed.
• Confidence: low with fewer than 8 answers, or with a range wider than 4,000 before 25 answers; medium from 3,000 to 4,000, or wider after 25 answers; high at 3,000 or narrower.
• Sharpen later: a new round continues from your earlier answers rather than starting over.

7. The level guide

The level label comes from the middle of your range, your number. Your gap list still starts at the bottom of your range. The level appears once confidence reaches medium; before that, the range is too wide to name a level:

The mapping is deliberately conservative. Published figures linking vocabulary size to CEFR levels vary widely, come mostly from learners of English, French and Spanish in classrooms, and count words in different ways. We chose thresholds at the cautious end. Treat the label as a rough guide, never as a certificate.

8. From your result to your gap list

• Every word you answered keeps your answer: solid, shaky or don’t know.
• For every other word, the list is split into blocks of 1,000. Blocks before the one that holds the bottom of your range count as probably known. That block and everything after it count as probably unknown, including the rest of your range. This errs toward showing you more words, not fewer.
• You can correct any word at any time. Topic lists and Anki exports use the same statuses.

9. Limitations

• Self-rating. People differ in how strict they are with themselves. Two learners with the same vocabulary can get different results.
• Recognition only. It measures whether you understand a word, not whether you can produce it.
• Cognates. For English speakers, words like hospital or idea are easy wins, so the estimate reads a little high.
• One meaning at a time. A lemma can have several senses. The second question catches some of this, but knowing one meaning still counts.
• Small samples. 10 to 40 answers can’t describe 10,000 words exactly. That is why you get a range.
• The list is the limit. Beyond 10,000 the test has nothing to measure.
• A smooth curve. Real knowledge can plateau, for example at 60–70% across several thousand words. The model’s S-shaped curve smooths that out, so the block chart shows a steady decline where your real pattern may be flatter.
• Slow faders get wider ranges. When knowledge fades gradually, 40 yes-or-no answers can’t pin the 80% point as tightly. The range says so rather than hiding it.

10. How we check it

Two kinds of checks, both run on the method as it stands today.

• Simulations. We generate thousands of virtual learners with known curves (some with a sharp drop, some fading slowly over thousands of words, some with word-by-word noise) and run the real test on them. With today’s model, the range contains the learner’s true 80% point 93–100% of the time for every kind we tested. The middle of the range typically lands within about 500 words, and up to about 1,000 for the slowest faders.
• Calibration runs. A learner rates 150 words at fixed ranks, 15 from each block of 1,000, with no adaptivity, so their real curve can be read off directly. The first one (October 2026) found the problem described under Changes. After the fix, that learner’s adaptive result (1,800–4,300) contained their calibrated point (about 3,600).
• One learner calibrated so far. One person is evidence, not proof. We’ll run more and report what they show here.

11. Changes

• 2026-10-16 — Slower fades allowed. A calibration run showed that the model, which then allowed only fairly sharp drop-offs, read a gradually fading learner about 1,600 words high. Three slower steepnesses were added, and positions now run to 18,000. Results got lower and ranges wider for many people.
• 2026-10-16 — Each word is chosen for what it tells us about your 80% point. Before, the app picked the word whose answer was hardest to predict. That asks near your 50% point, which can’t tell a sharp drop from a slow fade.
• 2026-10-16 — Stop suggestions also come after 25 and 35 answers, because honest ranges for slow faders often stay wider than 4,000.
• 2026-10-16 — Shaky counts as don’t know in the estimate. It used to count as half known, which simulations showed read several hundred words high.
• 2026-10-16 — The gap list starts at the bottom of your range, not its middle. The level briefly came from the bottom too, then returned to the middle once the model stopped reading high.
• 2026-10-14 — The adaptive assessment replaced the fixed test, which asked the same number of words in every block.

References

• Hacking, J. F., Rubio, F., & Tschirner, E. (2018). Vocabulary size, reading proficiency and curricular design: The case of college Chinese, Russian and Spanish.
• Milton, J. (2010). The development of vocabulary breadth across the CEFR levels. In Communicative Proficiency and Linguistic Development (EUROSLA Monographs 1).
• Milton, J., & Alexiou, T. (2009). Vocabulary size and the Common European Framework of Reference for Languages. In Vocabulary Studies in First and Second Language Acquisition.
• Nation, I. S. P. (2006). How large a vocabulary is needed for reading and listening? Canadian Modern Language Review, 63(1).
• Data: doozan/spanish_data (CC BY 4.0; frequency data CC BY-SA 4.0), Tatoeba (CC BY 2.0 FR).