TRUST

How the marking is checked

MedEngly marks practice writing and speaking with a two-pass AI assessor built against the published OET criteria. This page shows how we test it, what the latest test found, and what the marking cannot do. Only claims with retained evidence appear here.

LATEST CALIBRATION RESULT

HOW IT IS TESTED

1. Two assessors, every time. A first pass marks your letter or role play against each published criterion, quoting your own words as evidence before committing to any band. A second, independent critic pass verifies every score and must justify any change. You see the critic's result.
2. Stability before accuracy. The same letter marked five times must land within one band on every criterion. A marker that cannot repeat itself cannot be trusted, so this gate runs before anything else.
3. Calibration against references. The engine is scored against a private reference set of letters carrying deliberately planted, single-criterion faults, so every item has a known correct answer. Official published sample responses are used privately, during development only, to anchor the marker's band judgements; they are never shown to candidates. The bar we are working to, and have not reached yet: at least 90% of criterion scores within one band of the reference, confirmed against letters blind-graded by an experienced human assessor. That bar gates publishing an accuracy figure here and opening the public letter checker.
4. Every dispute is reviewed by a clinician. Each feedback screen has a dispute link. A human reviews within 48 hours, explains the decision, and the case joins the reference set so the same mistake is not made twice.

HONEST LIMITS

• This is practice feedback based on published OET criteria, not an official evaluation. Estimated grades are indicative, and only OET can grade your real exam.
• Speaking is marked from an automatic transcript. Intelligibility and fluency can only partially be judged from text, so those two bands are always labelled approximate.
• Reading and listening scores convert to the 0 to 500 scale approximately; the real test's conversion varies between sittings.
• Confidence is shown on every marking. When the two passes disagree, or a submission is unusual, the feedback says so instead of pretending certainty.
• The reference scores the engine is measured against are, so far, written in house. Until an independent assessor has graded the same letters blind, our accuracy testing cannot separate a marker that is too harsh from reference scores that are too generous. That independent grading is commissioned before any accuracy figure is published here.
• Content is currently the weakest criterion in testing, and the marking tends to be harder on it than the reference. If a Content band looks low to you, dispute it: every dispute is reviewed by a clinician and feeds back into the reference set.

MedEngly is built by a UK based IMG doctor who came through the international medical graduate pathway. Questions about methodology: support@medengly.com.