Back to home

Doucik Accuracy

Updated: July 2026

Doucik is artificial intelligence — and a certain error rate is simply part of that, just as no human tutor is mistake-free. We don't pretend otherwise: we measure the error rate in test lessons, publish it openly, and keep pushing it down with every improvement.

Minor slip

A clunky formula, a slower reply, a cosmetic glitch — it doesn't affect the learning.

~3 in 100 replies*

Wobble

Something doesn't work first time — a drawing or a graph. Doucik owns up and fixes it.

~2 in 100 replies*

Serious error

A reply fails completely, or a wrong result would be presented as correct. Rare — so always double-check important results.

fewer than 1 in 100*

Overall, more than 95 in 100 replies are completely fine. And the headline of the latest measurement (July 2026, 350 replies across ~19 test lessons): not once did we find a wrong mathematical result presented as correct. Even so, we plan for serious errors going forward — no AI is infallible, so always double-check important results.

Standard vs. Pro: the Pro plan uses a smarter model — it makes fewer mistakes on harder problems and has a steadier whiteboard and geometry. Its replies take a moment longer, because it thinks more.

* The numbers hold for clear communication — legible tasks, sharp photos, tidy writing on the whiteboard. With harder-to-read input (a blurry photo, a fragmentary task, illegible handwriting) there can be noticeably more mistakes. That's why a few simple habits pay off — see "How to get the most out of Doucik" below. Measured in test lessons simulating normal tutoring, updated July 2026.

How to get the most out of Doucik

A few simple habits that make mistakes even rarer:

  • Write legibly on the whiteboard — block letters, bigger digits. What Doucik can't read, it gets wrong.
  • Give tasks in parts — split a long problem into steps; have a complex construction drawn step by step (“first the triangle, then the height…”).
  • If it gets stuck, don't repeat the same thing — rephrase the task, or start a new conversation (a clean start = a clear head). Pushing the same question again and again only digs the error in deeper.
  • Don't agree with a result? Ask “check it again by substituting back” — Doucik verifies the working with a tool.
  • Verify important results — for tests and homework, always check by substituting back or in your textbook (a habit worth having even without AI!).
  • Drawing didn't work out? Type “try again” or “carry on” — Doucik notices the failure and fixes it; or switch to the other board (drawing ↔ geometry).
  • Report a mistake with the 🚩 button in the app — every report helps us make Doucik better.

What to do when a mistake happens

The app has a concrete button or feature for every situation:

  • A reply got stuck or never arrived → the “Try again” button by the notice, or the 🔁 Repeat quick action.
  • The result doesn't look right → the ✅ Check quick action — Doucik verifies the working with its calculator (or type “verify by substituting back”).
  • The drawing failed / the whiteboard looks empty → type “try again” or “carry on”; for a complex construction use 🗑 Clear board and have it drawn again step by step.
  • You don't understand the explanation → ❓ I don't understand or 🐢 Slower — it explains differently and more slowly, instead of repeating itself.
  • Doucik got “muddled” (answering off-topic, going in circles) → start a new conversation = a clean start; the old one stays saved in the sidebar.
  • Anything odd → the 🚩 Report a problem flag (every report helps us remove the error). And the “Need help?” advisor bubble helps right in the app.

How we measure and what's next

We measure the error rate in test lessons that simulate normal tutoring — a realistic hour with dozens of messages, photos of exercises, the whiteboard and graphs. Every Doucik reply is then classified into the categories above. We publish what a student actually sees and experiences in a lesson, and we round the numbers against ourselves("promise less, deliver more").

We update the numbers after every major improvement. AI models are also improving fast — each new generation means a lower error rate, and we roll them into Doucik as they arrive.

Changelog

  • 17 July 2026 — first published numbers (350 replies, ~19 test lessons): more than 95% of replies completely fine, 0 wrong results presented as correct. Deployed an automatic watchdog for stuck replies and a batch of whiteboard and graph fixes.

Doucik is a practice companion — it doesn't replace school teaching or checking results. The advisor bubble ("Need help?") is happy to answer accuracy questions right in the app, or email us at hello@doucik.com.