An AI tutor that always sounds confident is a marketing asset. An AI tutor that is always correct is a different thing entirely — and nobody has the second one yet. The honest version of "AI for education" includes acknowledging this and giving parents the data to weigh it themselves.
What we publish
On our /trust page, every month we publish per-(subject, level) pass rates from independent audits: a sample of real answers reviewed by a separate marker against the MOE marking scheme. Each row shows the subject, the level (Primary 1 through JC 2), the audit month, the sample size, and the pass rate as a percentage.
A "pass" means an answer the marker judges acceptable: factually correct, age-appropriate, and aligned with the marking conventions for that level. It is a stricter bar than "the AI gave an answer" and a weaker bar than "the AI gave the most elegant possible answer."
What we don't hide
Lower pass-rate cells are visible. If our Primary 4 Science audit came back at 78%, you can see 78%. Hiding the weak cells would be easy — most accuracy dashboards do — and it would also be dishonest. The weak cells are how we know what to fix next.
How the three-path system feeds into this
Behind every answer is one of three paths, marked with a verification badge. Bypass: we've answered a sufficiently similar question before and verified the cached answer. Grounded: we retrieved MOE-aligned source material and a separate auditor model concurred with the tutor's response. Ungrounded: no source material existed; the answer is the model's own and is clearly disclaimed.
Pass-rate audits sample across all three paths, weighted to actual usage. So the published number reflects the experience your child actually has, not a curated subset.
What this is not
It is not an accuracy guarantee. We will not offer money-back accuracy promises, because that creates incentives we don't want — to lower the bar, contest marker judgements, or focus engineering effort on the audit set instead of the real product. What we offer is the data, monthly, with methodology and sample sizes. You decide what threshold matters to your family.