AI tutors make mistakes. This is well-established and not in dispute. The interesting question is what a tutoring platform does about it. AL1 Tutor's answer is a three-path verification pipeline that makes the model's confidence — and lack of it — visible to the student rather than hiding it behind a polished tone.
Path 1: bypass (cached and verified)
When a student asks a question that is sufficiently similar — measured by vector embedding similarity — to a question we have already answered well, we serve the cached, verified answer. The badge on the answer reads "Verified." This is the highest-confidence path.
Bypass works because most questions a student actually asks at a given level have been asked many times before, in slight variations. The Mathematics question about three girls sharing 12 stickers in a 1:2:3 ratio has thousands of near-duplicates in our cache.
Path 2: grounded (retrieved + audited)
When no cached answer fits, we retrieve relevant MOE-aligned material — the worked solutions and reference content for that level and topic — and the tutor model drafts an answer using it. Before the student sees the answer, a separate auditor model checks it against the retrieved material independently.
If the auditor disagrees significantly, the answer is rewritten. If they concur, the student sees a "Reviewed" badge — meaning the answer is grounded in source material and a second model concurred, but it wasn't in our verified cache.
Path 3: ungrounded (model alone)
Sometimes — for novel questions, edge-case topics, or questions slightly outside the syllabus — no relevant source material exists. The model answers from its own knowledge, and the answer is clearly marked "Caution: ungrounded." We do not hide that this path exists.
The student is told, in plain language, to verify the working independently. This is not a failure mode we are embarrassed about; it is the honest representation of what the AI knows about that specific question.
Why three paths and not just "answer"
A single answer with no badge implies uniform confidence, which is wrong. A student who sees "Verified" on one answer and "Caution" on another correctly calibrates their trust in real time. They learn that AI is a tool with known failure modes, not a magic oracle. That calibration is part of the education, not a distraction from it.
You'll see the badge on every answer in the chat. Verified for bypassed cache hits, Reviewed for grounded answers, Caution for ungrounded. Tap the badge to see why it earned its rating.