What a score is made of
Each mark is one point of your score. The colour shows which criterion it belongs to.
The weights always add up to a hundred. If they do not, that set of criteria never reaches the site — the check is automatic. Below are the criteria for Everyday conversation. Here, asking again when you did not understand counts for more than vocabulary.
The moment you did not understand: did you say so, or nod and hope? This is the skill that decides whether you get the order right, not your vocabulary.
- Top mark, in the words it is given by
- Repaired without breaking the flow: confirmed back what was heard, flagged the exact gap, and the other person barely slowed down.
- Proof needed
- At least 1 checked quote from what you said, or the criterion comes back unscored
- Weight
- 30 of the hundred points here
100 points in all, 70 of them go to the other criteria
Asking when you're lost takes 30 points of 100
- This criterion
- Unassigned — always empty
What we do not promise
What the score still cannot tell you
That number came out of five checks. Not one of them answers the four questions below. A page that listed only the five would be selling the method instead of explaining it.
We prove you said it, not that it was right
The check proves the words are yours. Whether saying them was a good move, in that conversation, at that second, was decided by a model. That is why the wording behind every mark is published — so you can argue with it.
A score is not a language level
Difficulty describes the conversation — this one is pitched at B1 — and never the person having it. Nothing here certifies your level. A product that offered to would be making exactly the unprovable claim this page exists to replace.
Grading three times is still not certainty
The same conversation is graded three times and the middle result is kept. That is steadier than one go, and it is still a model reading a text. We measure separately how much the result wobbles on its own, and we refuse to draw any change smaller than that wobble. It is the honest way of saying we do not know.
None of this predicts the real day
Not your exam mark, not a visa decision, not whether you get the offer. We rehearse twelve minutes and hand back evidence about those twelve minutes. The rest depends on things this software cannot see.
After the score
You can still check a score months after you read it
When a conversation starts, we record five things: who you talked to, the situation, the criteria, the version of the scoring and the exact version of the model. That is why any score can be repeated and come out the same. Below is what happens when one of the five changes.
If the criteria change
Old scores are recalculated before you see them
Your conversation keeps the criteria it was graded by. A later change cannot quietly move a number you already read. If the criteria change, your history is recalculated first and only then drawn. Scores from two different versions are never joined by one line on a chart.
If the scoring changes
It does not switch on until it agrees with people
Before a new version of the scoring is switched on, it grades conversations that people have already graded. Its results are compared with theirs. Until they match, it stays off. The rule lives in the database itself, not in a document — that is the only kind of rule that survives a deadline.
When you ask
We explain any score in writing
Ask, and you get a written explanation of any automatic score: which criteria it was given by, which version of the scoring did it, and which of your own phrases it rests on. All of that is on the score already. This is your right to ask for it again, and it is written into our privacy page, not into a support reply template.
Where the facts come from
A score is built from three things. Only one of them is an opinion.
Two of the three are plain arithmetic over what you said. The third is the criteria, and they are published before you speak. So the one opinion here is the part you can read in advance and argue with afterwards.
Your own words
Under every line of the report
- Before you see a phrase, the server checks it word for word against the transcript of your conversation.
- A phrase that does not match is tried once more and then thrown away. We count how often that happens.
- So the model cannot reword you into agreeing with it.
The clock
Arithmetic, not opinion
- How much you spoke, how long your pauses were, how many filler words you used, how often you were interrupted.
- It is counted in code, so it comes out the same to the millisecond.
- It arrives a second after the conversation ends, before the score does.
The criteria
Published before you speak
- The criteria themselves, what each is worth as a percentage, and what each mark means.
- A criterion with nothing to back it stays unscored, instead of getting a polite middle mark.
- The total is worked out in code. We never ask the model for it.
Reporting
Think a score is wrong? A person reads it within four hours
Every conversation has a button to flag it. Every score has a link to dispute it. Here is what happens after you press one.
If it is about money
Money first, argument second
If you were charged for a conversation that did not run, we fix that before anyone discusses the score. We add a correcting entry instead of rewriting the old one, so the sums still add up. Your balance has always been the sum of those entries.
What helps
- Which conversation, and which criterion
- The phrase you think was misread
- What the score should be, in your view
You point at the score
Any time
The button sits under every criterion of every score. Two sentences are enough. The conversation it came from is attached for you.
A person reads it
Within 4 hours
With the transcript of your conversation and all three gradings open beside it. No template replies, and never the model checking itself.
The minutes come back
Same hour
If minutes were taken for a conversation that never ran, they come back before anyone argues about the score. Those are two separate questions.
You hear the outcome
Within 2 working days
In writing, saying what we checked and what changed. A new score is added next to the old one instead of replacing it, so you can still see the history.
FAQ
Six that a scoring page usually leaves to the fine print
Talk to a person
The only machines here are the people you talk to. Everything else is us, and we usually answer inside 2 hours.
Write to us