How a Langujet IELTS Band Score Estimate Is Built

Rubric-anchored scoring

The starting point is the public IELTS band descriptors, the same documents examiners are trained on. For Writing they describe Task Achievement or Response, Coherence and Cohesion, Lexical Resource and Grammatical Range and Accuracy at each band; for Speaking they describe Fluency and Coherence, Lexical Resource, Grammatical Range and Accuracy and Pronunciation. Each descriptor line is turned into features that can be measured. "Frequent error-free sentences" becomes the proportion of sentences with no detected error. "Uses a range of cohesive devices appropriately" becomes counts of reference, substitution and connectors and whether they are varied or repetitive. "Presents a clear position throughout" becomes whether a stance is detectable in the introduction, body and conclusion and whether it is consistent.

Anchoring to the rubric matters because it keeps the model honest about what it is scoring. A system that learned "what good essays look like" from examples would reward length and rare words; a system anchored to the descriptors rewards what the descriptors reward.

Why estimates are ranges, not points

Official IELTS scores are themselves the product of a human judgement with known variability: two examiners can differ by half a band on the same script. An automated estimate inherits that uncertainty and adds its own, because features approximate judgements rather than reproducing them. Reporting a single number such as "6.5" would imply a precision that does not exist. Langujet therefore presents each criterion estimate with its evidence and treats the overall figure as guidance, close enough to tell you which criterion is holding you back and whether you are trending towards your target, as the AI feedback page puts it. A difference of half a band between two attempts on different prompts is noise; a consistent half-band trend over six attempts is signal.

Calibration against examiner-marked samples

The model is checked against writing and speaking samples that trained examiners have scored, criterion by criterion. The method is straightforward: the model scores the same samples, and its criterion estimates are compared with the human scores to find where it is systematically generous or harsh, by band, by task type and by typical first-language error patterns. Adjustments follow where the comparison shows bias. The check is repeated as the model changes. We do not publish an accuracy percentage, because a single figure hides exactly the variation that matters: a model can be close on average and still be unreliable for one band or one task type, and the per-band, per-criterion picture is what the calibration is for.

What moves your estimate

• Task Response: answering every part of the question, stating a position early and keeping it. The fastest single gain for most candidates.
• Coherence: one idea per paragraph with a topic sentence; linking by reference ("this policy") rather than by stacking connectors.
• Lexical Resource: precise words in natural collocations; fewer repeated vague adjectives. Rare words used wrongly lower it.
• Grammar: the proportion of error-free sentences. Fixing your two most frequent error types moves it more than adding structures.
• Speaking fluency: fewer and shorter pauses while searching for language; pauses for thought are fine.
• Pronunciation: consistent sounds and word stress; accent is not scored.

Length beyond the minimum, rare vocabulary and memorised phrases do not move the estimate upward, and the last two can move it down, which is consistent with how examiners mark.

Where the estimate is least reliable

Three situations deserve extra caution. Very short answers give the model little to measure, so a Part 1 reply of two sentences or an essay far under the word count produces an estimate with wide uncertainty. Unusual but valid styles, such as deliberately short sentences or a first-person anecdote in a Task 2 essay, can be read as weaknesses by features tuned on typical candidates. And the top of the scale is harder than the middle: distinguishing a band 8 from a band 8.5 depends on judgements about precision and naturalness that features approximate only loosely. In each of these cases the per-criterion evidence is still useful; the overall figure is the part to treat with suspicion, and an examiner review is the way to settle it.

Reading your report

Read the four criterion estimates first and the overall figure last. Find the lowest criterion and open its evidence: the flagged sentences, the counted pauses, the repeated words. That evidence is the feedback; the number is a summary of it. Decide on one change for the next attempt. After several attempts, read the trend per criterion, which is the most reliable thing the report shows, because prompt-to-prompt noise averages out. The comparison of AI and human essay correction explains when to add an examiner review to verify what the trend suggests.

Practise with Langujet

The Langujet IELTS guide links each criterion to the practice that moves it, and every Writing and Speaking attempt on the platform produces the per-criterion report described here. Use IELTS writing correction for essays and IELTS speaking practice for recorded answers, and add an examiner-reviewed correction when you want an estimate verified.