Assessment research

Validity

A test is valid when it measures what it claims to measure. CELP tasks are built from situations candidates actually meet: a notice with opening hours, a platform-change announcement, an email to a landlord, a question about how you prefer to work.

We avoid tasks that reward test technique over English. A candidate who can read a timetable and write a clear complaint has demonstrated something real about their English; a candidate who has memorised a rubric has not.

Reliability

Reading and Listening are marked against a fixed key, so they are perfectly repeatable. The risk sits in Writing and Speaking, where two examiners could reasonably differ.

We manage that by sampling: a proportion of marked responses is second-marked blind, and disagreement beyond half a band is investigated rather than averaged away.

Fairness

Topics are chosen to be answerable by anyone, regardless of country, profession or background knowledge. A question that rewards familiarity with one country's institutions is testing general knowledge, not English.

Our invigilation analyses whether a face is present in frame, not whose face it is. It performs no identification, no demographic inference and no behavioural scoring.

Open questions

• How short can a speaking sample be before a band becomes unreliable?
• Does automatic marking of Writing reach examiner agreement on short functional tasks?
• What is the effect of a candidate's own device and connection on measured performance?