Methodology
How scoring works
The CELPIP-General test reports a separate level from 0 to 12 for each of its four components. There is no official overall score, so we don't invent one. Here is exactly how ScoreNorth arrives at each level, and where our estimates are weakest.
Listening and Reading
Each component has 38 scored questions. A correct answer scores 1; a wrong or blank answer scores 0, so always answer. Paragon publishes approximate raw-score charts whose bands overlap, because questions differ in difficulty. We report two things from one table:
- a point level, from a rule that splits each overlap near its midpoint, and
- a range: every level whose official band contains your raw score.
Worked example
30 correct in Listening falls inside the Level 7 band (27–31) and the Level 8 band (30–33). The point rule places 28–30 at Level 7, so we report Level 7, range 7–8, CLB 7.
| Level | Official band (raw / 38) | Our point rule |
|---|---|---|
| 10–12 | 35–38 | 35–38 → 10+ |
| 9 | 33–35 | 33–34 → 9 |
| 8 | 30–33 | 31–32 → 8 |
| 7 | 27–31 | 28–30 → 7 |
| 6 | 22–28 | 23–27 → 6 |
| 5 | 17–23 | 18–22 → 5 |
| 4 | 11–18 | 12–17 → 4 |
| 3 | 7–12 | 8–11 → 3 |
| M | 0–7 | 0–7 → M |
| Level | Official band (raw / 38) | Our point rule |
|---|---|---|
| 10–12 | 33–38 | 33–38 → 10+ |
| 9 | 31–33 | 31–32 → 9 |
| 8 | 28–31 | 28–30 → 8 |
| 7 | 24–28 | 25–27 → 7 |
| 6 | 19–25 | 20–24 → 6 |
| 5 | 15–20 | 16–19 → 5 |
| 4 | 10–16 | 11–15 → 4 |
| 3 | 8–11 | 8–10 → 3 |
| M | 0–7 | 0–7 → M |
Official bands: Paragon's published practice-test charts and the celpip.ca results page. Point rules: ScoreNorth's own inference. The charts can't separate Levels 10, 11 and 12, so we report “10+”. “M” has no CLB equivalent. Levels 3–12 equal CLB 3–12.
Practice sets with only some parts show your raw score per part, not a level: a level needs all 38 questions.
Writing and Speaking
Human raters score these components on four published dimensions each. Their exact rules for combining dimensions into a level aren't public, so ours are an approximation, and we say so on every result.
| Dimension | What it covers | Used in |
|---|---|---|
| Content / Coherence | Number and quality of ideas, organization, examples and supporting detail | Both |
| Vocabulary | Word choice, suitability, range, precision and accuracy | Both |
| Readability | Format and paragraphing, connectors, grammar, spelling and punctuation | Writing |
| Listenability | Rhythm, pronunciation and intonation, pauses and self-correction, grammar, sentence variety | Speaking |
| Task Fulfillment | Relevance, completeness, tone, and length or word count | Both |
- 1
Rate each dimension
Each dimension gets a level on the 0–12 scale, anchored to level descriptors written in our own words.
- 2
Combine into a task score
The four dimensions are averaged with equal weights.
- 3
Combine tasks
Writing averages its two tasks; Speaking averages its eight.
- 4
Report with a range
We round to a point level and show a range of one level either side (clipped to 0–12). We'll narrow it only where calibration shows we agree exactly with human raters at least 70% of the time at that level.
Floors: an empty response, or under 3 seconds of speech, scores 0 for that task. A response under 50 words (or a spoken answer under 20 words) is still rated, with Task Fulfillment capped at 4. Accent is never scored.
Basic estimate or AI grading
Always available, at no cost. A deterministic rubric checks measurable features such as length, task coverage, paragraphing, vocabulary range and sentence variety. It's consistent and explainable, but it can't judge the quality of an argument. Treat it as a rough guide.
When enabled, a language model rates each dimension and must quote evidence from your response for every rating. It runs within a daily allowance. Published research shows models like this agree with human raters reasonably well overall but tend to underrate strong writers, which is why we calibrate.
Every Writing and Speaking result is labelled with the grader that produced it.
Calibration and limits
Calibration metrics: not yet published
We'll publish agreement with human raters here (quadratic weighted kappa, and exact and adjacent agreement by level) once we have enough human-rated responses, along with a predicted-versus-actual chart from official results that users choose to share. Until then, Writing and Speaking results are marked uncalibrated.
- Our questions are original and written to match real-test difficulty, but they are not official test items.
- Real tests may include unscored trial questions; ours are all scored.
- Your result on test day depends on the day. Use ours to find weak spots, not to predict a number.