How AI scoring evaluates your CELPIP writing
Our AI marker scores a CELPIP Writing answer on Paragon’s four official dimensions — Content/Coherence, Vocabulary, Readability and Task Fulfillment — each from 1 to 12. Before it may give a level it fills in a ten-item evidence checklist out of 30, and its overall level must sit within one level of the band that total points to.
Official wording from Paragon’s CELPIP Results page, checked 4 November 2026. Our marker’s rules read from its source code the same day.
Try it on your own email: the free Writing Checker →
The four official rating dimensions
Paragon lists these categories and factors for Writing (its wording):
| Dimension | Factors |
|---|---|
| 1. Content/Coherence | Number of ideas · Quality of ideas · Organization of ideas · Examples and supporting details |
| 2. Vocabulary | Word choice · Suitable use of words and phrases · Range of words and phrases · Precision and accuracy |
| 3. Readability | Format and paragraphing · Connectors and transitions · Grammar and sentence structure · Spelling and punctuation |
| 4. Task Fulfillment | Relevance · Completeness · Tone · Word count |
Source: Paragon’s CELPIP Results page, “Writing Performance Standards”, checked 4 November 2026. Speaking uses the same structure with Listenability in place of Readability.
What the official test discloses — and what it doesn’t
Disclosed on the Results page:
- Writing is scored by qualified raters trained to apply standard scoring rubrics; each Writing performance is rated by four to six raters, independently, without seeing each other’s ratings.
- Each dimension is divided into five performance levels, and raters assign a level in each dimension from tangible evidence that matches the descriptors.
- When ratings disagree, a benchmark rater is assigned automatically, without knowledge of the first ratings.
- The component score is derived from the dimensional ratings and transformed into a CELPIP level by rules set in a standard-setting exercise with language experts.
Not disclosed there: the performance descriptors for each of the five levels, the transformation rules from dimension ratings to a level, any weighting between the four dimensions, and your dimension ratings themselves. The page says nothing about AI.
Source: Paragon’s CELPIP Results page, sections “Scoring the CELPIP Speaking and Writing Components”, “Rating Procedure” and “How is the final score determined?”, checked 4 November 2026.
How our AI marker applies them
These are the rules in the marker’s own instructions, in order:
- Evidence checklist first. Ten observable items, each 0 (no attempt), 1 (poor), 2 (adequate) or 3 (good), with a one-line note citing the answer. The Task 1 items come from a CELPIP preparation program’s answer key; Task 2 uses an adapted version.
- Four dimensions, 1–12 each, rated against Paragon’s published “At this level, I can…” level descriptors for Writing. The marker is told to judge what the writer can do, not to count errors.
- A holistic overall level, usually close to the dimension average — but an answer that fails the task cannot score high overall even if the language is strong.
- A guardrail, not a formula. Four officially levelled responses anchor the checklist total to a level (7/30 → 3, 9/30 → 4, 18/30 → 7, 27/30 → 9). If the overall level lands more than one level from the band the total points to, the marker must re-score the checklist, and say why if it still disagrees.
- Calibration rules: errors do not cap the level, and are not deducted twice; development and organisation outweigh polish; a clean but thin answer belongs at 5–7, never 9 on cleanliness alone.
- Feedback: a 2–3 sentence summary, one comment per dimension, 2–4 strengths, 2–4 prioritised improvements, 3–6 corrections of exact phrases, and the patterns that recur.
The Task 1 checklist
| Paragraphs | The email is organized in appropriate paragraphs. |
| Arguments | The email contains effective and detailed arguments. |
| Logical order | Ideas are organized logically. |
| Conventions | A greeting, an opener, a closer and a sign-off appropriate to the task. |
| Accuracy | Few grammar and spelling errors. |
| Transitions | Transition phrases and conjunctions improve the flow. |
| All points | Every point in the prompt is addressed. |
| Tone | The tone suits the task. |
| Vocabulary | The vocabulary suits the task. |
| Length | Between 150 and 200 words. |
Checklist total → level band
| Total /30 | 0–6 | 7 | 8 | 9 | 10–11 | 12 | 13–14 | 15–17 | 18–20 | 21–22 | 23–25 | 26 | 27 | 28 | 29 | 30 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Level | 1–2 | 3 | 3–4 | 4 | 4–5 | 5 | 5–6 | 6 | 7 | 7–8 | 8 | 8–9 | 9 | 9–10 | 10–11 | 11–12 |
Interpolated from the four anchors above, which is why the marker treats it as a guardrail.
An annotated Task 1 example, marked CLB 9
Prompt: You left your backpack on a city bus. Write an email to the transit company's lost-and-found office. This email was written by our own team as a test and marked by the AI on 18 August 2026; the text and the scores are the marker’s real output, unedited apart from the sign-off name.
Dear Lost and Found Office,
Conventions 3/3 — a proper greeting.
I am writing to report that I lost my backpack on one of your buses. I hope you can help me find it.
A clear opener that states the purpose. The marker suggested a more formal request: “I would greatly appreciate your assistance in locating it.”
I was travelling on bus number 22 yesterday afternoon, around 3:30 PM. I got on at Main Street Station and got off at Granville Street. I think I left my backpack under the seat near the back door.
Arguments 3/3 — concrete specifics: bus number, time, stops, where the bag was.
My backpack is black with a small blue logo on the front pocket. Inside, there is my laptop, my wallet with my ID card, and a notebook. These items are very important to me, so I hope someone found it.
Vocabulary 2/3 — clear but everyday (“very important”). Accuracy: one tense slip, “I hope someone found it” → “has found it”.
If you find my backpack, please call me at 604-555-1234 or send me an email at this address. I can also come to your office to pick it up if needed. Please let me know what time your office is open.
All points 3/3 — how to return it is covered. Transitions 2/3: paragraphs follow each other with few linking phrases.
Thank you very much for your help. I hope to hear from you soon.
A polite closer.
Best regards, [Name]
Sign-off. 163 words in all — Length 3/3.
| Dimension | Level | Marker’s comment (shortened) |
|---|---|---|
| Content/Coherence | CLB 9 | Relevant, well developed, logically sequenced; each paragraph has one purpose. To go higher: one more explicit link between paragraphs. |
| Vocabulary | CLB 8 | Accurate and appropriate, but mostly common words. |
| Readability | CLB 9 | Clean grammar, punctuation and spelling; some complex structures, a wider range would lift it. |
| Task Fulfillment | CLB 9 | All three required points covered, email conventions followed, length within range, tone matched. |
| Overall | CLB 9 | Checklist 28/30 (Transitions and Vocabulary 2/3, every other item 3/3) — the 28 band is 9–10, so 9 is inside the guardrail. |
What would lift it: the marker’s first two improvements were explicit transitions between paragraphs (“As for the contents,” “Regarding how to reach me,”) and more precise words in place of “very important” and “help”.
The limits
- It is one AI pass, not four to six trained raters with a benchmark rater, and it cannot use Paragon’s unpublished performance descriptors or transformation rules.
- The checklist-to-level bands rest on only four officially levelled responses.
- The result is an estimate for practice. It is not an official CELPIP score, and Paragon Testing Enterprises has no connection to Celpip.me.
- The same answer can come back a level apart on a different day or model; read the comments and corrections, not only the number.
Frequently asked questions
How does AI scoring evaluate CELPIP Writing?
Ours scores the answer on Paragon's four Writing dimensions — Content/Coherence, Vocabulary, Readability and Task Fulfillment — each on the 1–12 CELPIP scale against the published level descriptors. Before it assigns any level it must fill in a ten-item evidence checklist (0–3 each, out of 30), and its overall level has to land within one level of the band that checklist total points to.
Is the official CELPIP Writing test marked by AI?
Paragon's CELPIP Results page, checked 4 November 2026, says Writing is scored by qualified, trained raters: each Writing performance is rated by four to six raters working independently, and a benchmark rater is assigned when their ratings disagree. That page says nothing about AI scoring.
Is an AI CELPIP Writing score the same as an official score?
No. It is an estimate from a practice tool. The official level comes from several human raters and a transformation set by Paragon's standard-setting experts, which Paragon does not publish. Use an AI score to see what to fix, not as a prediction you can rely on for an application.
Does the AI take marks off for every grammar mistake?
No. Our marker is told explicitly that errors do not cap the level: an email organised into real paragraphs, with developed points, connectors and the right tone, can reach CLB 9 with recurring small slips. A clean but thin answer is placed at 5–7, not 9.
Run the same marker on your own answer
Paste a Task 1 email or Task 2 survey answer: a CLB estimate, a reason for each criterion, two strengths and two fixes. Free, no account.
Open the Writing Checker →