She will not give you the answer.
On anything graded, Ara will not give you a hint until you have committed an answer. The prompt is “Try it first — then I can help”, not a lock — but it does not open early, because a hint taken before an attempt replaces the retrieval that was the point.
lib/lesson/hintGate.ts
She cannot give you credit you did not earn.
Marking is generous about your wording and strict about your substance. “The ice melts” and “melting occurs” score the same. The model that judges paraphrase is handed the authored credit points and nothing else, so it cannot invent credit you did not earn, and in almost every case it can only move your mark up. The exception is disclosed rather than hidden: when the two markers disagree completely — the scheme says you earned everything and the model says you earned nothing — neither is trusted, and the answer is capped at half marks instead of one of them deciding.
lib/lesson/gradeLeniency.ts
She cannot remember something that was never recorded.
When Ara refers back to something you got wrong three weeks ago, she is reading it from a stored record of that lesson rather than reconstructing it. She formats and filters what was actually saved, and can say nothing that was not — which is the difference between a tutor who remembers you and one who is fluent about it.
lib/lesson/learnerModel.ts
What she is not
Ara does not read your emotions, watch you through a camera, or judge your energy — this page claimed all three and none of it was built. She arrives at the point where you got something wrong, which is a smaller promise and the only one worth making.
Every answer you give — inside a lesson, on a card, in a mock paper — goes through one function. There is one record of what you know, not one per feature, which is why the app cannot congratulate you on one screen and warn you on another.