Writing and speaking evaluation: official criteria, double scoring
A text goes through two separate evaluations and the result is averaged. The criteria follow the telc form for the level; feedback in Turkish or English.
The multiple-choice sections grade themselves; the real question is writing and speaking. In the real exam trained assessors score these two sections with a level-specific form. A candidate studying at home gets either no feedback at all or nothing beyond 'nice work'.
On Level Exams writing and speaking answers are scored by AI with a one-to-one equivalent of the level's official assessment criteria. The same answer goes through two separate evaluations; if the scores diverge noticeably, a third tie-break evaluation steps in. The result is not a single number but a score per criterion, concrete corrections and an explanation.
Below: how scoring works in the real exam, our method, and the number of criteria per level.
How writing and speaking are scored in the real exam
The writing text is scored against the criteria on the level's assessment form: covering the given points (Leitpunkte), communicative effect, accuracy of vocabulary and grammar, suitability for the text type. At the higher levels criteria are scored in discrete bands (A/B/C/D) and converted to the final score with a central multiplier.
Speaking is scored task by task; at the lower levels with separate criteria per task (introduction, spelling, question and answer, request), at the higher levels with fluency, language, pronunciation and task fulfilment per task.
Assessors are calibrated to the expectation of the level: at A1 short, simple sentences are not a flaw but a full-score answer. This is the most critical rule of scoring, and it is exactly where most 'AI corrections' go wrong.
How Goethe and ÖSD score
Writing and speaking are scored by examiners at all three providers against that level's official criteria — a criterion-by-criterion reading, not a free impression.
The criteria sets differ. In Goethe's modular exams writing and speaking are each scored out of their own 100 with no compensation between them; at telc both belong to the same exam and the written/oral group thresholds apply.
What they share comes down to four axes: content (did you cover every given point), fitting the text type, linguistic accuracy, and being understood. Those four meet you whichever provider you sit.
The AI scoring here uses telc's level criteria. If you are sitting Goethe or ÖSD the feedback still helps, because the same four axes are measured; the point scale, however, is not identical to that provider's.
How Level Exams evaluates
The criteria and the level-expectation note are defined in the format file and go into the evaluation prompt verbatim; the engine does not 'know' which level it is scoring, it only applies what the form asks for.
Criteria from the telc form
Each level's writing and speaking criteria were defined as a one-to-one equivalent of the official assessment form. Every row you see in the score table is a row in the real exam too.
The same answer twice, then the average
Your answer goes through two independent evaluations and is averaged. If the two diverge noticeably a third tie-break is called. There is no arbitrary single score.
Level expectation is enforced
Language models tend to score lower-level texts too harshly. Each level's form carries a note on what a full-score answer looks like at that level, and it goes into the evaluation prompt verbatim.
Form-filling tasks do not go to the AI
Single-correct-answer tasks such as form filling at A1/A2 are scored by rules, with a defined spelling tolerance. AI is used only for free production.
Speaking is transcribed first
Your recording is converted to text by speech recognition, then scored with the task criteria. At levels whose official criteria include pronunciation it counts; at the others it is additional feedback only.
Concrete corrections and explanations
Next to the score for every criterion: what was missing and how the sentence should have read, explained in your language. Not 'could be better' but 'the second Leitpunkt was not covered at all'.
Evaluation criteria by level
The number of criteria for writing and speaking, and how many times an answer is scored, come live from the format file.
Why double evaluation
A single evaluation can be inconsistent
The same text can get a different score when read twice; that is true of human assessors too. Averaging reduces the chance element.
The tie-break catches divergence
If the two scores are far apart, the answer really is borderline; a third evaluation settles it, and you see that in the report.
Criterion-based feedback is actionable
'70 points' teaches nothing. 'Communicative effect full, grammar half' tells you what to work on.
How to get the most out of the feedback
- 1
Look at the criterion column, not the total
The criterion where you scored low tells you what to practise; the total only tells you where you stand.
- 2
Rewrite the corrections
Write the corrected sentences from the report again in your own hand; reading does not teach, writing does.
- 3
Try the same task type again a week later
When the same kind of task appears in another mock, apply the earlier feedback and watch the criterion score rise.
- 4
Write at your level, not above it
Trying to build long, complex sentences at A1 raises the risk of errors. The evaluation is calibrated to the level; simple and correct beats complex and flawed.
Frequently asked questions
›How close is the AI score to the real exam score?
Because the criteria and level calibration are taken from the official form and two evaluations are averaged, it is a close estimate in direction and magnitude; but it is not an official score and does not amount to a certificate.
›Which model is used?
Anthropic's Claude for evaluation and Whisper for speech recognition. Even if the model changes, the method stays the same because the criteria and calibration come from the form.
›What language is the feedback in?
Your interface language: Turkish or English. Exam content and instructions are always in German.
›How long does scoring take?
Evaluation starts the moment you finish the section; writing takes a few seconds, speaking a little longer including transcription. It appears once the results report is ready.
›Can I dispute an evaluation?
This is a preparation tool, not an official exam. If you find a score questionable, report it via support; recurring deviations are reviewed to improve calibration.
Have one of your texts scored
Write the writing section of the free mock exam and see the score per criterion and the corrections with your own eyes.
Start the free mock exam