How Attest English works
Fifteen minutes of everyday English across all four skills, scored against the CEFR with human moderation. Every result is backed by a full evidence record: the recording, the transcript, the scores and the reasons. This page sets out exactly how a result is reached, so you can hold us to it.
The standards we work to
Council of Europe framework
Every score reports against the Common European Framework of Reference, the international standard for describing language ability.
Good practice in language testing
Our design follows the published principles of the international language testing associations: ALTE's minimum standards for quality, and the ILTA and EALTA guidelines for good practice. Validity, reliability, fairness, transparency and the right of appeal.
Human oversight
No candidate can fail by machine alone. Borderline, flagged and appealed results are decided by people.
UK data protection law
Identity data, recordings and automated decision-making are handled under UK data protection law, including your right to make representations and to ask for human intervention.
One sitting, eight stages
The assessment runs hands-free in your browser. Each stage begins on its own, you speak or type, and it moves on. Here is the full sequence, and the evidence each stage gives us.
Practice
WARM UPNot scoredThree warm-up items: say your name, repeat a sentence, answer a question
Nothing here is scored. It exists so the real tasks measure English, not nerves.
Spoken interview
SPEAKINGThree questions about you, your day and your plans · up to 45 seconds each
Free speaking across past, present and future forms.
Read aloud
READING · SPEAKINGTwo short everyday notices, read at a natural speed
Pronunciation, pace and reading fluency, measured against the printed text.
Listen and repeat
LISTENING · SPEAKINGSix sentences, each heard once and repeated exactly
Listening accuracy and spoken control, from simple sentences to long ones.
Scenario
SPEAKINGOne everyday situation to describe and resolve · up to 60 seconds
Connected speech: describing what is happening and explaining what you would do.
Listening questions
LISTENINGA voicemail or announcement, played twice, then three spoken questions shown on screen from the second play
Comprehension of natural spoken English: times, places and instructions. Playing the message twice tests listening, not memory.
Reading comprehension
READING · WRITINGA short everyday text with three typed answers
Understanding written English and answering clearly in writing.
Short message
WRITINGA typed message built from given facts · 60 to 120 words in eight minutes
Writing: a clear, accurate, complete short message under a timer.
Questions, notices, sentences, scenarios, messages and texts are drawn from banks of variants, so no two assessments need contain the same material.
Every item is about buses, appointments, messages and plans. Specialist knowledge is never tested, because the question is whether you can converse.
Any phone or laptop, no test centre
The assessment runs in the browser on any phone or laptop with a camera and microphone. No appointment, no test centre, nothing to install. The same proctoring, identity checks and scoring apply wherever it is taken, which is the model modern remote-proctored testing uses.
We know who sat the assessment
A certificate is only worth what its identity evidence can prove. Four checks run on every assessment, and all of them are stated on screen before you consent.
A photograph, taken at assessment
Before the assessment begins you take a framed photograph of yourself. That photograph appears on your certificate, so the certificate shows the person who sat the assessment.
Photo ID, checked by AI
You hold your passport, driving licence or residence permit up to the camera. The photo is checked the moment you confirm it, so an unreadable document is caught while you can still retake it. At scoring, an AI vision check compares the face on the document to your photograph and reads the printed name against the name you registered with. Anything unclear goes to a person.
Your name, in your voice
In the first practice item you say your full name aloud. That recording is kept with the evidence record, anchoring a voice sample to the identity evidence.
Continuous recording and monitoring
Camera and microphone stay on from consent to the end, behind a visible recording banner. Monitoring images are captured every thirty seconds during the spoken tasks, the camera is checked for a second person every few seconds, and the recordings are checked for a second voice.
On the record from consent to completion
Recording begins when you consent and ends when the assessment ends, behind a recording banner that stays on screen the whole time. Monitoring images are captured every thirty seconds during the spoken tasks. Nothing is recorded covertly, and nothing is recorded before you consent.
The recordings, transcripts, photographs and monitoring images live in a private store, never on the open web, and are read through short-lived signed links. They are seen by the organisation that sent you your assessment link and by our reviewers, never by the public. A certificate check at verify.attestenglish.com confirms a certificate; it exposes none of the evidence behind it.
Three steps, and the machine never has the last word
Scoring separates what should be measured from what should be judged, and keeps the final arithmetic fixed and published. The same answers always produce the same score.
Transcribe and measure
Every spoken answer is transcribed with word-level timings by a pinned speech-to-text model, with a second engine as fallback, and the engine used is recorded on the result. Code, not AI judgement, then computes the deterministic measures: speech rate, pauses, and word-for-word accuracy against the reference text for the read-aloud and repetition tasks.
An AI examiner marks each task
An AI examiner scores each task from 0 to 5 against CEFR descriptors, one score per criterion, and must quote the candidate's own words as evidence for every score it gives. Written tasks add spelling and punctuation. The examiner estimates a CEFR band per task, and it never decides pass or fail: its output has no such field.
Why the AI never decides pass or fail
The examiner's output has no pass or fail field to fill in. Its criterion scores are inputs to the fixed arithmetic below, doubtful cases go to a person, and the exact transcription and examiner model versions are stored in every result record, so any result can be traced to the models that produced it.
Fixed, published arithmetic
The task scores are combined by fixed weights into a composite from 0 to 100. The weights, the cut scores and the review bands are constants in code, stored with every result, and published here. For read aloud and repetition, measured accuracy carries 60 per cent of the task score and the examiner 40 per cent; on the listening and reading tasks, comprehension counts twice in the examiner's mean.
Every assessment is set against a target level, stated on the result and the certificate. A B2 target uses a cut score of 65 with a human review band of 58 to 72; a B1 target uses a cut score of 50 with a review band of 44 to 58. The reported CEFR estimate is independent of the target: the same composite always maps to the same estimated level.
The safeguards
The arithmetic decides the easy cases. Everything doubtful goes to a person, and these rules are enforced in code, not policy.
Borderline goes to a person
Any composite inside the target level's review band (58 to 72 for a B2 target, 44 to 58 for B1) is decided by a human assessor. The band deliberately brackets the cut score, so no borderline candidate is ever decided by the machine alone.
A collapsed criterion goes to a person
A passing composite can hide one collapsed skill. If any single criterion on any task scores 1 or below, a would-be pass is held for human review instead.
Contradictory evidence goes to a person
Near-native reading aloud beside collapsed free speaking usually means a fluent speaker giving minimal answers, not a weak speaker. When the evidence disagrees with itself, a person decides.
Poor audio is never failed
If a recording is missing, too short or too unclear to score fairly, the machine is not trusted in either direction: the result is held for a person, or the attempt restarts at no cost. Audio problems are never marked as a fail.
Identity concerns go to a person
If the photographs, the document or the name do not clearly match, a human confirms who sat the assessment before any outcome stands.
Every decision stores its full working
Each result keeps its weights, per-task scores, formulas, flags and reasons, alongside the recordings and transcripts, so any result can be re-derived and re-checked line by line.
A level you can place
Results are reported against the Common European Framework of Reference, the international standard for language ability. A B2 result is comparable to Functional Skills English Level 1 and ESOL Level 1 (RQF), the benchmark used for UK licence-linked security training.
Every certificate carries the photograph taken at the assessment, the holder's name, the date, the overall CEFR level and the level in each skill, under a unique certificate number. Anyone can verify a certificate free of charge at verify.attestenglish.com.
Your rights in every result
Language, never length
The examiner is instructed to measure the quality of your English, not the amount of it. A short answer in fluent, natural, accurate English is strong evidence of ability and is scored as such.
Human review on request
Any candidate can request a human review of their outcome at any time. A person re-marks the recording, and the human decision stands.
Retakes after 48 hours
One attempt per person per 48 hours. If the day goes badly, you can sit the assessment again two days later.
The evidence record
Every result keeps its recording, transcript, per-criterion scores with quoted evidence, and the arithmetic that produced the outcome.
Where calibration stands
A new assessment has to earn trust, so here is the honest position. Results are monitored continuously. Borderline results and anything flagged are marked by a human assessor, and those human judgements are kept alongside the automated scores. That growing set of double-marked assessments is our calibration data.
As it grows, the cut scores will be anchored against double-marked human examiner judgements, and a dated technical report will be published on this page. Until then, the human review band stays deliberately wide, so people decide the cases the machine has not yet earned the right to.
- The task weights, cut score and review band, exactly as implemented today.
- Any change to them, dated, before it applies to new assessments.
- A dated technical report, once the calibration set supports one.