Surface what’s already there
Every exchange opens by asking the student to attempt or recall. Nothing is handed over before a real try — even a failed attempt makes the concept stick harder.
One distributes, the other ingests, and between them sits a syllabus, a grade — and now two private AIs. Homework used to carry a little signal about how a student thinks. It doesn’t anymore. Raya replaces all of that with one shared thread.
Three surfaces, one profile underneath. Nothing to configure, nothing to keep up to date.
Raya climbs one rung at a time and never starts at the top. The goal isn’t a correct answer in the next thirty seconds — it’s a student who can do it again next week, alone.
This isn’t a setting a school switches on. Raya is built so it cannot hand over a finished answer — and its feedback targets the method, never the person, per Carol Dweck’s work: “this one trips a lot of people up”, never “you’re gifted”.
The Kernel doesn’t store a grade. Every concept carries a four-part vector — and when a signal degrades it names the failure, so Raya can change tactic on the very next turn.
Raya, the classroom version of Raya, the graded homework — not one of them keeps its own copy of a student. They all write into the same profile and read it back before they answer, while the conversation is still going.
k_raw · k_effectiveMastery of that one concept, raw and adjusted, recomputed after each attempt.
v_scoreThe observed learning rate, so a slow climb reads differently from a full stall.
p_scoreResistance to forgetting: a concept secured on Monday that quietly slips by Friday gets caught.
m_scoreWhether struggle reads as information or as proof of failure. Tracked per student, not per concept.
Concepts are stored as a structure, not as a syllabus. So when something breaks, the Kernel doesn’t stop at the exercise that broke — it walks back along whatever that concept rests on, until it reaches ground that holds. That ground is often in a subject nobody thought to check.
Your LMS records what was done, on the days work is due. A tutor helps in the moment, and then the moment is over. Neither one keeps what a student actually understood — so on an ordinary Tuesday, with nothing due and no session running, there is nothing to read.
Everything in this category runs on bought frontier models. So does this, from the same suppliers. The model is the floor — it is not what anyone is choosing between. The record above it is.
And that record is narrow on purpose. A teacher sees that derivatives are fragile. A teacher does not see what was said — which is less than the gradebook they already have knows about the same student.
Three ways in — a solo student, a whole school, or a bespoke deployment. Pick your lane.
Free to start, no card. The student begins; the Kernel does the rest.
Solo, or with your class. Nothing to install.