Most platforms marketed as adaptive adjust difficulty the way a thermostat adjusts temperature. If the student gets too many wrong, it turns the heat down. If they get too many right, it turns the heat up. The mechanism is simple and it sounds reasonable until you realize it treats all wrong answers as equivalent — which they are not.
A student who gets a problem wrong because they've never seen that concept is in a completely different situation from a student who gets it wrong because they have a specific incorrect belief about how the concept works. The first student needs instruction. The second student needs their incorrect belief surfaced and confronted before any instruction will land. A difficulty slider cannot distinguish these cases. It just turns the heat down for both of them and calls it adaptive.
The ACDE doesn't adjust a difficulty number. It reads what the student actually understands and targets the exact concept at the exact tier that sits at their current growth edge — then gets more precise with every session as the profile updates.
Before every session, the Adaptive Concept Difficulty Engine reads the student's Dynamic Learning Profile — their schema floor on every concept branch, the specific wrong beliefs that have been flagged, the calibration state from last session, the language ceiling. It uses all of that to select which concept to target, at which tier, with which type of problem.
During the session, it monitors accuracy pattern in real time. If accuracy collapses below the productive zone, it flags for immediate recalibration — not more encouragement, not more practice at the same level. Recalibration.
After every session, it updates the profile. Schema floor advances if the student demonstrated it. Wrong belief flags update if a misconception was addressed. Calibration state adjusts based on actual performance. The next session starts from a more accurate picture than the last one.
The system compounds. By session five, it knows more about how this student learns than most teachers discover in a semester — not because it's smarter but because it's tracking the right things and never stops.
In the current implementation, tutors review every problem before a student sees it, and every significant change to the student's profile is reviewed by the tutor rather than made automatically. This is intentional. An automated system that updates student profiles without human oversight can make mistakes that compound — a wrong schema flag incorrectly resolved, a schema floor incorrectly advanced — and those mistakes affect every subsequent session.
The engine proposes. The tutor validates. As the system accumulates more data and demonstrates reliability, the automation layer expands. But the human review gate stays in place at the points where mistakes would cause the most damage.