# Apple Global School case study: validated with 150+ students | SwaVid | SwaVid How SwaVid validated its learning profiles with 150+ students at Apple Global School: 365 test submissions, 191 feedback responses, 85% teacher-confirmed accuracy, and the limits of the evidence. Canonical: https://swavid.com/case-studies/apple-global-school Source: https://swavid.com/case-studies/apple-global-school # Validated in the classroom, not the lab. ## Why a real school, not a spreadsheet. ## The headline result: 85 % confirmed. ## The cohort, in data. ## What the feedback changed. ## What this evidence cannot tell you. ### Can a family read it? ### Does it match the child? ### Does it point somewhere? ### Why diagnosis-first is the only correct architecture for AI tutors ### The tutor should know what to teach before it starts talking. We put SwaVid's learning profiles in front of 150+ students at Apple Global School and asked the teachers who knew them best to grade every one. 85 % were confirmed accurate. This report shows the method, the exhibits, and exactly what the evidence cannot claim. The principle under test Map the learner. Teach the missing idea. Check again. Adapt. A chat response answers the question in front of you. A learning system must also notice the idea underneath it. This engagement tested whether that principle survives contact with a real classroom. Figures from the evidence review dated 24 July 2026 . Submission counts include repeated attempts and must not be read as unique-student counts. SwaVid is built on a strong opinion: an AI tutor that answers before it understands the learner is a faster way to be wrong. Opinions are cheap. So we ran a classroom validation with Apple Global School and asked three uncomfortable questions. Is the learning profile clear enough to anchor a real conversation at home, without a teacher translating it? Does the teacher who sees the student every day recognise the child in the generated profile? Does the profile end in a concrete next step, something to teach next, rather than a label? Teachers reviewed each generated profile against the child they teach. In 85 % of reviews, the profile was confirmed as an accurate reading of the student. Share of profile reviews where the teacher said the report matched the child they see in class. 365 submissions from 200 learners, 64 completed deep reports. Real product data is messy, and the mess is information: the June wave is the school-wide rollout, the long spring tail is voluntary re-attempts. Amber marks the cohort's weakest trait: structured, step-by-step thinking Before the report avg 4.47 · n = 191 After the report avg 4.66 · n = 32 Feedback is only real if it costs you something. Across 191 responses, five demands recurred: plainer language, better pacing, a more useful report, no fixed labels, and one concrete next step. Each is now in the product. Report language Rewritten for family conversations: plain, non-diagnostic, free of fixed labels. Pacing Shorter steps and fewer concepts per screen, shaped by completion and feedback timing. The handoff Every profile now ends in a next step: identify the prerequisite, teach it, check again. What we publish Internal scoring fields stay private. Public pages carry the evidence limits next to the findings. “The product direction is strongest when the profile connects to a concrete next step: identify a prerequisite, teach it, and check again.” Evidence review, 24 July 2026 We publish the limits in the same report as the findings because that is what makes the findings worth reading. When the evidence grows, this page grows with it: review date, refreshed consent and cohort checks, and all. The learning science behind the principle this validation put under pressure: Bloom's 2-sigma problem and three decades of knowledge-tracing research. That is why SwaVid starts with a learning map, not an empty chat box. See the diagnosis-first flow this validation sharpened. - This is a classroom validation of profile quality, not an efficacy study. It does not measure grade improvement. - The cohort comes from one school and does not represent all students. - Submission counts include repeated attempts and are not unique-student counts. - Post-report feedback (n = 32) is a small, self-selecting sample. We treat it as direction, not proof. - No consent metadata existed for publishing individual responses, so no names, emails, answers, or quotations appear here. - The learning profile is not a clinical, diagnostic, or standardized assessment. ## Key Links - [White paper Why diagnosis-first is the only correct architecture for AI tutors The learning science behind the principle this validation put under pressure: Bloom's 2-sigma problem and three decades of knowledge-tracing research. Read the white paper](https://swavid.com/research/why-diagnosis-first) - [See the learning map](https://swavid.com/learning-debt-identifier)