The Adaptive Concept Difficulty Engine (ACDE) is the automated calibration component of the Lacefield Adaptive Learning System. It reads the student's Dynamic Learning Profile, selects the appropriate concept node and difficulty tier for each session, monitors performance data in real time, and updates the DLP after every session. It also accumulates adjacent concept signals and proposes tier map refinements when signal frequency warrants.
The ACDE is not a difficulty slider. It is a schema-aware targeting system that knows the difference between a student who got a problem wrong because the concept is new and a student who got it wrong because their model is corrupted.
1. Read DLP: Load schema_floor per branch, wrong_schema_flags, fluency_bottlenecks, calibration_state, language_ceiling. This is the complete picture of where the student is.
2. Select target concept: Based on schema_floor and calibration_state, identify the highest-tier concept that is at the growth edge — sound enough to build on, not so far ahead that the student lacks prerequisites.
3. Generate problems: Pass target concept, active misconception flags, student SCP, and calibration target accuracy to Claude API. Receive problems with misconception-tagged distractors.
4. Monitor accuracy: Track per-problem accuracy during session. Flag destructive frustration signals if accuracy collapses below 50% or engagement pattern shifts.
5. Log adjacent signals: Any error that doesn't map to a known misconception ID is logged as an adjacent_concept_signal on the nearest node.
6. Update DLP: Post-session, update schema_floor if performance demonstrates tier advancement. Update wrong_schema_flags resolution status. Recalculate calibration_state direction.
7. Propose tier map refinements: When adjacent_concept_signal frequency on a node crosses threshold (default: 3 instances across 2+ students), flag for expert review.
Tutor performs ACDE functions manually using the DLP and session log. Problem generation via Claude API with tutor review before delivery. Calibration updated by tutor after each session.
Session logs fed to Claude API for pattern matching. Automated calibration proposals with human approval. Adjacent concept signal aggregation across students. Live data integration for tier map refinement queue.
Real-time DLP updates. Automated per-session calibration without human approval. Destructive frustration detection triggers automatic difficulty drop. Full tier map refinement pipeline with expert review gate.
ACDE is not item response theory. It does not model difficulty as a property of individual problems independent of student state. Difficulty in this system is relational — the same problem can be growth-edge for one student and mastery reinforcement for another depending on their schema floor. The ACDE reads the student's DLP to determine the appropriate targeting, not a generic difficulty parameter.
ACDE is not a recommendation engine. It does not optimize for engagement, retention, or session time. It optimizes for schema development at the correct difficulty tier. These are different objectives that produce different systems.
No tier map change goes live without human review. The ACDE proposes. A qualified human validates. This constraint is non-negotiable in MVP 1, 2, and 3. Automated tier map changes without expert review risk propagating a misidentified adjacent concept signal into the dependency structure — which would corrupt instruction for every student whose path runs through that node.