How Is a French Exam Paper Actually Graded?
Two independent markers, blind, and agreement measured at 0.71 across 6,569 TCF writing scripts.
Your writing paper is not read by one person. The TCF candidate handbook, April 2026 edition, describes two independent markings carried out blind. And across 6,569 real scripts, a study presented at TALN in July 2024 puts the agreement between those two markers at 0.71. Here is what that means for you.
Who marks a French writing paper, and how?
The TCF writing test has three tasks. According to the candidate handbook from France Éducation international, you get 60 minutes for the TCF tout public, the TCF Québec and the TCF Canada, and 30 minutes for the TCF IRN. Expected lengths run from 60 to 120 words for task 1, 120 to 150 for task 2, 120 to 180 for task 3; the IRN version asks for 30 to 60 words, then 40 to 90.
Then comes the marking, and here the handbook is blunt: "Les épreuves d'expression écrite (EE) sont évaluées 2 fois de manière totalement indépendante et en double aveugle", that is, each paper is assessed twice, fully independently and blind. Each marker gives each of the three tasks a level ranging from "A1 non atteint", below A1, up to C2. Six levels are therefore recorded for a single paper, and then "a calculation rule taking the 6 levels into account produces a final mark and a final level". The candidate receives a mark out of 20 translated into a CEFR level.
One practical consequence slips past most candidates: the handbook sets the level at "A1 non atteint" if a task is not written with the required number of words. Counting your words is not a detail, it is a condition. If you are still choosing between the two exam families, our article on TCF or DELF compares what each is used for, and the one on registration dates gives the deadlines to plan for.
Do two markers give the same level to the same paper?
Nobody asks that question, and it has been measured. The TCFLE-8 study, presented at JEP-TALN-RECITAL in July 2024, brings together 6,569 TCF writing scripts, 581,333 words, from candidates declaring eight different usual languages.
On that corpus, agreement between the two markers, measured as quadratic weighted kappa, is 0.71. Agreement between the markers and the level finally retained for the script rises to 0.84. The spread by level shows what they read: 689 scripts at A1, 1,375 at A2, 1,466 at B1, 1,427 at B2, 1,127 at C1 and 485 at C2.
That 0.71 explains the rest of the procedure: the handbook states that where the two markers differ substantially, a third marking is carried out as a matter of course. Read alone, that line looks like a formality; read next to the measured figure, it becomes a necessary part. Two trained professionals agree often, not always.
English speakers are the largest single group in this corpus: 917 of the 6,569 scripts come from candidates who declared English as their usual language, ahead of Arabic at 906 and Spanish at 904. So the figures above are not a distant average for you, they describe a pool in which your language profile is the best represented one.
Can a machine mark a French paper?
Partly, and the handbook says so plainly. Some computer-based writing papers may receive a first marking from the FIDELIA tool; the second marking stays human, and a candidate who refuses that assessment can ask for the paper format. Since October 2024 the handbook describes a hybrid double marking, again with a third human marking where the two diverge sharply.
The study's figures place those tools precisely. CamemBERT reaches 0.88 quadratic weighted kappa against 0.93 for the human raters. But it lands on the exact level only 57 % of the time, against 76 %. Its adjacent accuracy, the share of scripts placed at the right level or one level away, reaches 0.98 against 0.99. The two other models tested, XGBoost and a logistic regression, trail at 0.79 and 0.69 kappa.
The machine is rarely very wrong, but it is often slightly wrong. That is exactly the error profile a second human reading catches.
Is the DELF marked the same way?
No, and confusing the two wastes preparation time. The DELF and DALF examination rules, published in Bulletin officiel no. 31 of 27 August 2009, set a total mark out of 100, each paper marked out of 25, out of 50 for the DALF C2. The pass mark is 50 out of 100, and above all there is a minimum mark required per paper of 5 out of 25, raised to 10 out of 50 for the DALF C2.
The consequence is sharp: at the DELF, a total of 50 is not enough if a single paper falls below 5 out of 25. At the TCF there is no threshold and no failure, only a level. And the DELF regulation describes neither double marking nor blind marking: the procedure above belongs to the TCF, not to French exams in general. Our articles on the DELF B1 and on preparing for the DELF B2 set out the papers at each level.
What do these figures not tell you?
Four caveats, and they matter. First, 0.71 quadratic weighted kappa is strong agreement, not disagreement: nothing here supports the claim that marking is arbitrary. Second, the 6,569 are scripts, not candidates, and the spread describes that corpus, not the TCF population. Third, one author of the study, Vincent Folny, works for France Éducation international, which publishes the handbook: the two sources are not fully independent, even though one is a peer-reviewed conference paper and the other an official document. Finally, the corpus holds TCF scripts only, so 0.71 does not transfer to the DELF, which has its own juries.
Three moves follow. Count the words in each task before handing in, since a task outside the range scores "A1 non atteint". Treat the three tasks as three separate assessments, since each marker scores three. And at the DELF, look at your lowest mark first rather than your total, because that is the one that can rule you out.
Counting words and reading a regulation takes a few minutes on your own. Knowing whether what you write holds the level you are aiming at takes someone who reads it the way a marker would. At Frenchee, classes are live with qualified native teachers, one to one or in small groups of 3 to 5, with no subscription, and our level test gives you a first marker.
