How we grade writing and speaking

AI grades written essays and spoken answers in the trainer — against the same criteria used to assess language on international B2-level exams. Below is what the score is made up of, what you see in the result, and where this kind of grading has its limits.

The main thing up front: the Promovat score is a practice guide, not an official ANCE grade. It shows what's already working and what's worth working on. Your actual exam result may differ — the score doesn't claim to predict it.

What it's based on

ANCE doesn't publish its own scale for grading writing and speaking — like most citizenship exams, it relies on the CEFR descriptors (CECRL in Romanian): descriptions of what a person can do at each level. The target level is B2. Those same descriptors are the basis of our grading.

To make the score predictable and defensible, we built it from what international B2-level language exams agree on: a similar set of criteria, scales with anchor descriptions, and the rule that «at B2, it matters more that speech is clear than perfectly error-free». We used the public rubrics of Cambridge B2 First and Goethe-Zertifikat B2 as a model.

Promovat is not affiliated with or endorsed by Cambridge, the Goethe-Institut, the Council of Europe, or ANCE. We reference their public materials as a model of exam practice, nothing more.

• CEFR — qualitative aspects of spoken language (Council of Europe, table 3) — the official level descriptors, including B2
• Cambridge B2 First — writing assessment scales — the public Cambridge English rubric
• Goethe-Zertifikat B2 — exam regulations — the Goethe-Institut assessment criteria
• ANCE — citizenship exam — the official exam page and a link to the CECRL descriptors

The criteria we use

The score is made up of several criteria, each on its own scale. Writing has four, a spoken answer has five (fluency and pronunciation are added). The final score is calculated from the criteria, not set by gut feel.

Written essay — 4 criteria

Spoken answer — 5 criteria

At B2, the odd mistake is fine as long as it doesn't get in the way of the meaning. So grading looks not at how many errors there are, but at whether they affect understanding.

What the result looks like

After grading you see not just a score, but everything that went into it:

• A score from 0 to 10 and a profile for each criterion — where you're already good and what to work on.
• An error breakdown: each one with a word-for-word quote from your answer, a correction, and a short explanation.
• For a spoken task — a transcript: how the system heard your answer.
• A short summary and specific tips on what to work on before your next attempt.

We hold one rule firmly: a comment only appears alongside a word-for-word quote from your answer. No quote, no comment. That makes the grading itself easy to check and saves you from arguing with a «score by impression».

Honest limits

AI grading is a strong training tool, but it has limits, and we don't hide them:

• AI can miss an error or give a debatable call on a borderline case — especially in grammar and pronunciation. That's why the score doesn't hang on one or two findings.
• The score is approximate. It's a guide for practice, not a prediction of your exam grade.
• If a score seems unfair, the product will have a «disagree with the grading» button. Those signals help us tune it.

We deliberately show not just the score but everything it rests on — quotes, transcript, criteria. A score you can check yourself earns more trust than a number out of nowhere.