Conjuga.pro is an app to practise Spanish verb conjugation. I made it to learn myself: it is a web app (PWA) on Google Play with Supabase/PostgreSQL. Every answer is saved. This analysis uses data from 3 May – 4 November 2026: 6,390 answers on 158 active days.
n = 1 (me). Observational data, except the randomized test below.
Case 1
My accuracy stopped growing. More of the same practice did not help much.
1 · The problem: a plateau
With the same mix of tenses every month, my accuracy grew from ~65% in May to 74–78% in June and July. After that it stayed at 76–81%.
“Accents forgiven”: if the answer is wrong only in a written accent, it counts as correct. Same mix of tenses every month. November has only 4 days.
2 · Diagnosis: what the data says
The data did not support the idea “I don't practise enough”.
69%
of the 4,004 forms I saw came up only once.
Random questions spread a little practice over thousands of verb × tense × person combinations.
59%
correct when I met the form again after a mistake.
n = 1,149, median gap 1.1 days. You can't compare this directly with 77% the first time: the forms I get wrong are harder anyway.
×1.22
higher odds of a correct answer each time I double my practice in a tense.
95% CI 1.17–1.27. Real, but small. Practice and total study time grow together, so the model can't separate them.
Accuracy in the first third and the last third of my answers in each tense, from weakest to strongest. Hover or tap to see the last two months. The imperfect subjunctive dropped the most (−4 points); two other tenses dipped by 1–2 points.
3 · The fix: Smart training
Launched on 7 November 2026. Instead of random questions, every session is built from my history:
40% reviews
20% weak spots
40% new forms
Mistake→back in 3–5 questions→tomorrow→longer gaps
Reviews use spaced repetition; weak spots are the tenses and forms with the most mistakes. It has automated tests and goes to a test site first, then to production. Random mode is still in the settings.
4 · The experiment
RUNNINGAnalysis plan and prediction written before the start
Does repeating a mistake in the same session help more than repeating it only the next day? This is the main part of Smart training.
First mistake on a form→coin flip 50/50→repeat now (3–5 questions later)ornext day only→correct at the next encounter?
Design
Within-person randomized experiment (one learner), from 9 November 2026. The unit is a mistake on a form. I don't see which group a mistake is in. Both groups get the same review the next day.
Main metric
Correct answer the next time the form comes up, on another day.
Analysis
Difference in proportions with a 95% CI, two-proportion z-test, α = 0.05.
Sample & power
3 weeks, about 150 mistakes per group: enough to see a difference of about 15 points (for example 59% → 74%) with about 80% power, but not a smaller one.
Control
One session a week in random mode with the same settings. I don't compare daily accuracy between modes: Smart mode gives harder questions on purpose.
Limits
n = 1; it tests the repeat step, not all of Smart training. Answers from one learner are not fully independent, so the z-test is an approximation, and the result says nothing about other users.
Results
29–30 November 2026.
The data suggests the metric stopped not because of how much I practised, but because of which questions I got. So I changed the product instead of “trying harder”, and a randomized test planned in advance checks the effect.
Case 2
Check the data before you analyse it
Before any conclusions, I checked the data itself. I use the same rule at work with LLM tags: check the labels before you act on them.
Graded every answer again
Check I graded all 6,390 answers again against the correct forms.
Found27 saved as wrong but actually correct (for example reflexive -se forms) and 23 saved as correct but actually wrong.
Fixed Case 1 uses the corrected grades.
Checked the reference data
Check I wrote my own rule-based conjugator and compared it with the open verb database: 11,180 forms.
Found2 errors in the original database.
Fixed Corrected; this check now runs before every release.
A bug found in the data
Check I compared the app's own stats with the raw logs.
Found On days when I reached the goal of 30 correct answers, the database had only 29: the last answer was not saved.