Home/Methodology

SCORING & FEEDBACK

Practice feedback should tell you what to improve next.

UpskillPrep uses CEFR-aligned estimated practice feedback to help learners identify strengths and weaknesses. It is designed for preparation and learning, not for issuing an official Cambridge English result or CEFR qualification.

What an UpskillPrep result means

Practice results are estimates generated from the learner's activity inside UpskillPrep. They are intended to support preparation decisions such as which skill to practise next, whether an answer needs more detail, or whether timing and delivery need attention.

They do not replace an official Cambridge English test report and should not be presented to an employer, school or other organisation as an official qualification.

Estimated CEFR guidance

Where the product shows A1, A2 or B1, the level is an estimated CEFR-aligned preparation signal. CEFR describes language ability through proficiency descriptors; it does not define one universal raw-percentage formula that every test can use.

UpskillPrep therefore keeps the scoring methods for objective and productive skills separate, then combines comparable internal CEFR evidence for the four-skill estimate. The internal evidence scale is an UpskillPrep calculation and is not the Cambridge English Scale.

Reading and Listening

Reading and Listening objective answers are checked against the assessment answer data, not judged by an AI model. Raw accuracy is mapped to provisional A1–B1 preparation boundaries. Those defaults are calibration settings rather than universal CEFR cut scores and can be replaced by set-specific boundaries as benchmark data grows.

Writing feedback

Writing is evaluated across task achievement, grammar, vocabulary, coherence and language range. The final A1–B1 estimate is calibrated by product rules rather than accepting a model's suggested CEFR label directly.

Task coverage, sufficient development and derived communicative-achievement evidence are also considered so length or grammatical accuracy alone cannot produce a B1 estimate. Learners should still consider purpose, reader, tone and register when revising a response.

Speaking feedback

Speaking evaluation considers task achievement, fluency, grammar, vocabulary and coherence. Timed transcription can also provide delivery evidence such as pace, pauses, fillers and continuity.

For assessed read-aloud tasks, the app now compares the recognized wording with the supplied reference text as a conservative intelligibility check. This is useful evidence that speech was understandable enough to be recognized, but it is not a phoneme-level pronunciation, accent, stress or intonation score. Regional accents are not penalized simply for being different.

Recorded one-way responses also cannot fully measure interactive communication such as turn-taking or negotiating meaning. The Speaking result should therefore be read as an estimate from the speaking tasks actually assessed, not as a claim about every CEFR spoken-interaction descriptor.

How the four-skill estimate is combined

The overall level no longer averages four raw percentages as though a Reading percentage and a rubric-derived Writing percentage meant exactly the same thing. Each completed skill is first placed into an internal evidence band consistent with its A1, A2 or B1 estimate, and the four comparable evidence values are then combined.

This keeps the overall calculation stable while preserving the distinction between objective accuracy and productive-language rubrics. The internal evidence number is not presented as an official external scale.

Confidence and retries

Automated language evaluation includes a confidence signal. A valid result with limited evidence is treated as an automated estimate rather than being sent into a required human-review queue. Learners should repeat the task when the available evidence is weak. Technical scoring failures remain retryable instead of being converted into a language judgement.

Limitations and validation

  • Automated language feedback can be imperfect and should be used as preparation guidance.
  • Audio quality, response length and available evidence can affect Speaking feedback confidence.
  • Read-aloud recognition similarity is intelligibility evidence, not full pronunciation assessment.
  • A single attempt is not enough to describe a learner's full English ability.
  • Default score boundaries remain provisional until they are calibrated against a sufficiently large benchmark sample.
  • UpskillPrep mocks use independent practice content and are not official Cambridge English papers.

How calibration improves over time

The scoring engine records versioned calibration information so future benchmark studies can compare automated estimates with independently assessed learner samples without silently changing old reports. As enough learner data becomes available, Reading and Listening sets can receive set-specific boundaries and productive-skill thresholds can be checked against benchmark ratings.

How to use feedback well

Look for patterns across attempts rather than reacting to one number. Choose one concrete improvement, repeat or revise the task, then check whether the same issue appears again. Over time, that is more useful than chasing a higher score without changing the response.