Between Sept 2025 and Mar 2026 we collected 10,247 Speaking recordings (Part 1 + 2 + 3) from paid users in Vietnam. Each recording was graded twice: 1. Britwise AI (GPT-4o + Whisper, scored against the Cambridge Public Band Descriptors). 2. A panel of 4 certified IELTS Speaking examiners (avg 9 years' experience). Both sides graded blind. We compared the overall band and each of the 4 criteria (Fluency, Lexical, Grammar, Pronunciation).

• 95% of Britwise scores fell within ±0.5 band of the examiner consensus. • 71% landed within ±0.5 of all 4 criteria simultaneously. • Disagreement was symmetric (we under- and over-grade in equal measure — no systematic bias). • Where Britwise and the examiner panel disagreed by >0.5 band, internal review found the AI was correct 41% of the time (the human panel itself sometimes splits).

We tracked 312 paid users who completed at least 30 days of daily practice (≥3 sessions/week). On entry, their mean Speaking band was 5.4. After 30 days the mean was 6.6. Median gain: +1.0; 90th percentile: +2.0; 10th percentile: +0.5. Comparable studies for traditional 1-1 tutoring (£30/h equivalent) show ≈ +0.5 band in the same time frame for the same starting level.

• Not a Cambridge Assessment endorsement. We aim to mirror their rubric but we are independent. • Not a guarantee. Individual outcomes vary. Our +0.5 band in 30 days refund policy underwrites the average outcome — not every case. • Not the entire test. Listening, Reading and Writing are scored with separate methodologies (Listening/Reading are auto-marked; Writing uses a similar 4-criteria GPT-4o panel).

Every quarter we re-run the calibration with a fresh batch of 500+ recordings. If the AI drifts beyond our 95% ±0.5 threshold, we recalibrate prompts and rubric weighting until it returns. Want to audit the data? Email research@britwise.info with your institution and we'll share the de-identified calibration set under NDA.