Schema tracing is the core diagnostic method of the Lacefield Adaptive Learning System. Every architectural decision in the system — the DLP structure, the problem generation prompts, the calibration engine, the tier maps — follows from it.
Knowledge tracing (BKT, PFA, DKT, and their variants) tracks which skills a student has been exposed to and whether their performance history predicts future success. It is a surface-level model — the answer is the data point.
Schema tracing treats the answer as a symptom. The data point is the mental model that produced the answer. A student who consistently gets linear equations correct may be operating from a corrupted model of what an equals sign means — one that happens to produce correct answers on standard problem formats but collapses on novel presentations. Knowledge tracing cannot detect this. Schema tracing is designed specifically to detect it.
| Dimension | Knowledge tracing | Schema tracing |
|---|---|---|
| Primary data point | Answer correctness | Mental model structure |
| Failure mode detected | Absent knowledge | Absent knowledge + corrupted models |
| Wrong schema detection | Not possible — correct answers mask corrupted models | Core function — misconception IDs flag specific corruptions |
| Response to failure | More exposure to the concept | Surface and confront the specific misconception |
| Personalization unit | Skill mastery probability | Schema floor + active misconception flags per branch |
| Self-correction mechanism | Accuracy history updates priors | Adjacent concept signals propose tier map revisions |
A student can produce correct answers at Tier 3 while operating from a Tier 1 misconception. Knowledge tracing marks this as mastery. Schema tracing flags it as wrong schema with a misconception ID. These produce opposite interventions.
Schema tracing distinguishes three states for any concept node, each requiring a different system response:
Treating wrong schema as absent schema is the most common and most damaging implementation error in adaptive system design. The interventions are not similar. They are opposite.
The Dynamic Learning Profile stores schema state at the structural level required for schema tracing. The critical constraint: wrong_schema_flags and schema_floor are separate objects and cannot be collapsed.
{
"schema_floor": {
"algebra_equality": {
"tier": 2,
"confidence": 0.85,
"last_assessed": "2026-05-19"
},
"linear_functions": {
"tier": 1,
"confidence": 0.70,
"last_assessed": "2026-05-19"
}
},
"wrong_schema_flags": [
{
"branch": "algebra_equality",
"corrupted_tier": 1,
"apparent_tier": 3,
"misconception_id": "MIS-EQ-01",
"misconception": "Equals sign as answer indicator, not relational balance",
"evidence": "Student confused when equation presented as 5 = 2x + 1",
"resolution_status": "active"
}
]
}A student can have a schema floor of Tier 2 in one branch and Tier 1 in another simultaneously. A wrong schema flag can show apparent Tier 3 performance built on a Tier 1 corruption. Any data structure that averages these or collapses them into a single mastery score destroys the precision that schema tracing requires.
Every concept node in the tier maps carries a misconception set — specific wrong beliefs that produce predictable error patterns. These are not generic wrong answers. They are the exact incorrect models that students build, named and catalogued.
Misconception IDs follow the format MIS-[BRANCH]-[NUMBER]. They are used in three ways: to generate diagnostically useful distractors (wrong answer options that represent specific misconceptions, not random wrong numbers), to flag active wrong-schema patterns in the DLP, and to select diagnostic probes that surface a specific misconception.
A problem generation call that receives MIS-EQ-01 as an active flag generates a problem that will reveal whether the student reads equals as a balance relationship or as an answer indicator — not a generic equation problem. This is the mechanism that makes schema tracing operational rather than theoretical.
When a student fails at a concept for a reason that doesn't map to any existing misconception ID, the system logs it as an adjacent concept signal. These signals indicate that the tier map is missing a prerequisite. Accumulated across students and sessions, they are the raw data that allows the tier maps to self-correct over time.
Adjacent concept signals must be logged from the first session. They cannot be reconstructed retroactively. They are the most valuable long-term data the system produces.