From Magic to Mathematics: Let Assessment Define the Starting State

From Magical Mysteries to Mathematical Meaning · Article 6 of 7

The transition from magical mystery role play to curricular design began with a structural equivalence. In a good wizard-detective mystery, the player finishes in a state they could not legitimately occupy at the beginning. They know who caused the magical event, how it was accomplished, or why the apparent impossibility was possible, and they possess evidence that justifies that conclusion. In good instruction, the learner should likewise finish in a state they could not legitimately occupy at the beginning. They should understand something, distinguish something, explain something, or perform something, and there should be evidence that the change is real.

Both are state-transition problems. The designer knows the intended endpoint before the participant reaches it. That makes curriculum amenable to the same inverse architecture that emerged from the mystery system.

Education already has backward design, of course. Teachers routinely begin with standards, objectives, or desired outcomes and then plan instruction. The contribution here is not the slogan "start with the end." It is the attempt to make the intermediate epistemic structure explicit enough that an LLM can reason over it as persistent project state.

A topic label is not a terminal state. "Teach derivatives" names content. Even "Students will understand derivatives" is too vague to determine what must precede it. A more useful target represents what learners should be able to discriminate, explain, predict, justify, or perform.

Suppose a quantity is modeled by a function M(t), where t is time. A meaningful terminal state might require students to interpret M'(t) as the instantaneous rate of change of the modeled quantity with respect to time; explain the units of M'(t); use the sign of M'(t) to reason about whether M is increasing or decreasing; and connect that interpretation across a formula, a graph, a table, and the context of the model. That is specific enough to backsolve.

Immediately before students can make that interpretation, what must already be available to them? They may need a stable concept of average rate of change over an interval and an understanding that the units of a quotient carry contextual meaning. Before instantaneous rate is usable, they may need to coordinate secant slope with change in the modeled quantities and then understand what happens as the time interval is made smaller. If a common misconception is that a derivative is merely an algebraic procedure or "the slope of the graph" without reference to the variables and units, the trajectory may require contrasting examples that make the contextual meaning indispensable. Each transition can be represented as a prerequisite relation rather than left as an intuition in a lesson plan.

This resembles the mystery's evidentiary trajectory for a reason. The mystery designer asks what observation would justify the next inference, what world fact would produce that observation, and what earlier condition would make the world fact true. The curriculum designer can ask what experience would justify the next conceptual move, what representation or task would produce that experience, and what prerequisite understanding must already be present for the learner to use it.

But curriculum introduces a difference that becomes decisive. The mystery designer usually creates the solver's initial state. The teacher does not create the students' initial states. The students arrive with them.

That means the inverse solve should not stop at a hypothetical learner profile if actual assessment evidence is available. Assessment can estimate the present learner state and therefore determine how far backward the curriculum must be constructed.

This gives assessment two distinct positions in the architecture. At the far end, assessment supplies evidence that the target state has been reached. At the near end, assessment supplies evidence about the state from which instruction must begin. The same instrument need not serve both purposes, but the state model should distinguish both functions.

Imagine that the target requires prerequisite concepts A, B, and C. A designer working from a generic grade-level expectation might assume all three and begin at the next transition. A short diagnostic may reveal something else. One learner has A and B but not C. Another has B and C but a misconception about A. A third can execute a procedure associated with all three but cannot explain the relation among them. Those differences are not administrative details. They define the actual boundary conditions of the design problem.

The backsolve continues until the trajectory reaches states the learner can genuinely enter. If the learner lacks C, the trajectory needs a bridge to C or must begin before C. If the learner has C but holds a misconception that makes the next inference unavailable, the bridge must address that misconception. If the learner can perform the procedure without the conceptual structure required by the final objective, the design cannot treat procedural success as sufficient evidence of readiness.

The persistent state representation makes this more operational. The curriculum project can contain objects for target capabilities, prerequisite concepts, misconceptions, examples, counterexamples, tasks, assessments, and observed learner evidence. Relations can say that one capability requires another, that an example supports a distinction, that a diagnostic response indicates a misconception, or that an assessment item tests a particular target.

Once those relations exist, a change in assessment evidence can trigger a recomputation of the instructional trajectory. This is the same leverage that the mystery system gained when world facts were externalized. The model does not merely remember that a student "struggled." It can update a represented learner state, traverse the dependencies that lead to the objective, and identify which transitions are no longer safely assumed.

This also changes how adaptive instruction should be understood. Adaptation is not simply generating an easier worksheet or another explanation when a student gets something wrong. The more precise question is: what state transition is currently inaccessible, and what prerequisite or evidentiary bridge would make it accessible?

Suppose a student fails to distinguish mass from weight. If the later objective requires reasoning about gravitational force, the response is not necessarily to add more practice on the current lesson. The system can inspect which later transitions depend on that distinction, identify the earliest point at which the learner's model diverges from the required path, and construct an intervention at that point. The intervention is justified by the dependency structure, not merely by the fact that the latest answer was incorrect.

This is also why assessment should be represented semantically rather than only numerically. A score of 70 percent is often too lossy for inverse design. Two students with the same score may have different patterns of understanding and therefore require different bridges. What matters is which prerequisite states are present, absent, unstable, or contradicted by the evidence.

The mystery analogy clarifies the problem. Knowing that a detective wizard has found "70 percent of the clues" tells us almost nothing about whether the solution is accessible. We need to know which clues were found, which propositions they support, and which inference remains unavailable. Curriculum is similar. Aggregate performance can summarize, but it cannot substitute for the topology of prerequisite access.

The human remains responsible for defining the objective and judging the instructional meaning of the evidence. The LLM's advantage is that it can operate over a larger represented structure than a teacher can conveniently hold in working memory. It can compare assessment evidence against prerequisite graphs, propose candidate bridges, test those bridges against the target state, and update derivative lesson materials without losing the canonical instructional objectives.

The same mutation discipline from the mystery application matters here. A newly inferred learner state should be treated as a proposal grounded in evidence, not as an unquestionable fact about the student. Assessment is noisy. A wrong answer can result from misunderstanding, inattention, language, or an ambiguous item. The system should preserve provenance: which observation supports which inference about learner state, with what confidence or qualification.

The design therefore remains formal-semantic rather than purely symbolic. Formal relations tell the system where dependencies run. Semantic inference helps interpret what a student's response means and what kind of bridge might address it. The teacher can inspect, correct, or override those interpretations before they become canonical state.

For one learner, this gives a clean inverse-design loop. Define the terminal state. Backsolve its prerequisites. Use assessment to estimate the learner's actual current state. Continue the backsolve until the trajectory reaches that state, adding bridges where necessary. Then materialize and execute the trajectory forward, updating the learner state as new evidence arrives.

A classroom adds one more level. The teacher is rarely designing twenty entirely separate curricula. Twenty students must usually participate in a shared instructional trajectory. Their assessed states are heterogeneous, but the class still needs a common path toward common objectives.

That means the next state we need is not simply another learner profile. It is a class-level initial condition: a collective assessed boundary that tells us how far the design must backsolve so the trajectory is accessible to everyone.