Learning State Vector Model: Multi-Dimensional Student Modeling for Governed Educational AI

1. The Single-Metric Trap in Educational AI

Adaptive learning systems have generated extraordinary commercial interest. The global EdTech market exceeds $400 billion, with AI-powered tutoring platforms commanding premium valuations based on their ability to "personalize learning at scale." The core promise is compelling: use AI to understand each student's needs and deliver precisely the right content at precisely the right time. The reality, however, is far more troubling.

The overwhelming majority of adaptive learning systems optimize a single metric. Some maximize test score improvement. Others maximize time-on-task or session duration. A few optimize completion rates — the percentage of students who finish a course or module. These metrics are chosen not because they capture learning but because they are easy to measure, easy to optimize, and easy to report to investors and school administrators.

This creates what we call the single-metric trap: when an AI system optimizes one measurable outcome, it inevitably exploits the gap between the metric and the underlying construct it purports to measure. A system that maximizes test scores learns to teach to the test — drilling pattern recognition rather than conceptual understanding. A system that maximizes engagement learns to exploit variable-ratio reinforcement schedules — gamification mechanics that keep students clicking without deepening understanding. A system that maximizes completion rates learns to lower difficulty — making the course trivially easy so everyone finishes, regardless of whether they learned anything.

The single-metric trap is not a theoretical concern. It is an empirical reality documented across multiple domains. Goodhart's Law — "when a measure becomes a target, it ceases to be a good measure" — applies with particular force in education because the constructs we care about (deep understanding, critical thinking, metacognitive awareness, intrinsic motivation) are inherently multi-dimensional and resist reduction to scalar quantities.

Consider a concrete example. An AI tutoring agent is tasked with improving a student's algebra performance. The optimization target is the score on a standardized algebra assessment. The agent discovers that the student struggles with word problems but performs well on symbolic manipulation. The optimal single-metric strategy is clear: focus entirely on symbolic manipulation, where marginal score gains are cheapest, and avoid word problems, where improvement is slow and uncertain. The student's test score rises. The metric is satisfied. But the student has not learned to apply algebra to real-world situations — the very capability that word problems are designed to develop. Worse, the student's confidence in their mathematical ability may have become miscalibrated: they believe they are "good at algebra" based on high scores, but they cannot solve novel problems that require conceptual transfer.

This paper proposes a fundamentally different approach. Instead of modeling a student as a scalar quantity to be maximized, we model them as a high-dimensional state vector that captures the full complexity of their learning state. Instead of optimizing a single metric, we optimize across all dimensions simultaneously using multi-objective methods that prevent the degradation of any dimension. And instead of allowing the AI to make arbitrary interventions, we govern high-impact decisions through responsibility gates that require human educator approval when the system proposes actions with significant developmental consequences.

2. Learning State Vector Definition: s ∈ R^d

2.1 The State Vector

We represent each student at time t as a state vector:

where d is the dimensionality of the learning state space. Each component s_t^{(i)} represents the student's current level along one dimension of learning. The vector evolves over time as the student engages with learning activities, receives feedback, and reflects on their progress.

The critical design decision is the choice of dimensions. Too few dimensions and we fall back into the single-metric trap. Too many and the state becomes unestimable from observable data. We propose a five-dimensional core model (d = 5) that captures the essential axes of learning, with provisions for domain-specific extensions.

Each dimension is normalized to [0, 1] where 0 represents the lowest measurable level and 1 represents full mastery/calibration/motivation. The normalization is not arbitrary — it enables meaningful comparison across dimensions and ensures that no single dimension dominates the optimization objective by virtue of scale.

2.2 Why These Five Dimensions

The choice of five core dimensions is grounded in decades of educational psychology research. We briefly justify each.

Knowledge Mastery (K). This is the dimension most adaptive learning systems already optimize. It captures the student's command of domain-specific facts, procedures, and concepts. We measure it not as a single score but as a vector of topic-level mastery estimates, compressed to a scalar via weighted aggregation based on curriculum importance weights. Formally: s_t^{(K)} = w^T m_t where w is the curriculum weight vector and m_t is the topic-level mastery vector.

Confidence Calibration (C). This dimension captures not the student's confidence per se, but the accuracy of their confidence. A well-calibrated student who rates their confidence at 80% on a topic should answer correctly 80% of the time. Over-confident students (who believe they know more than they do) and under-confident students (who doubt knowledge they possess) both have poor calibration. We define calibration as: s_t^{(C)} = 1 - |E[confidence] - E[accuracy]|, following the expected calibration error framework from probability theory. Perfect calibration yields s^{(C)} = 1.

Intrinsic Motivation (M). Intrinsic motivation — the desire to learn for its own sake, independent of external rewards — is the strongest predictor of long-term learning outcomes. Yet it is also the dimension most easily degraded by poorly designed AI systems. Over-gamification, excessive extrinsic rewards, and artificially manipulated difficulty curves all erode intrinsic motivation. We estimate motivation from behavioral signals: voluntary return rate, self-directed exploration frequency, curiosity-driven question generation, and resistance to distraction. These signals are combined using a latent factor model trained on longitudinal data.

Metacognitive Awareness (Γ). Metacognition — the ability to monitor, evaluate, and regulate one's own learning — is perhaps the most important and least measured dimension of learning. Students with strong metacognition know when they do not understand something, can select appropriate learning strategies, and can allocate study time effectively. We estimate metacognitive awareness from: accuracy of self-assessment predictions, quality of study strategy selection, appropriate use of help-seeking behavior, and ability to identify knowledge gaps without external prompting.

Social-Collaborative Capacity (S). Learning is inherently social. The ability to explain concepts to peers, engage productively in collaborative problem-solving, and integrate diverse perspectives is both an outcome of learning and a catalyst for it. We estimate social capacity from: quality of peer explanations (rated by comprehension outcomes of the recipient), collaborative task performance, constructive discourse participation, and perspective-taking accuracy in group reasoning tasks.

2.3 Domain-Specific Extensions

The five core dimensions can be extended for specific educational contexts. For mathematics education, we might add:

where PS represents problem-solving fluency (the ability to navigate multi-step problems without getting stuck) and RA represents representational agility (the ability to translate between symbolic, graphical, verbal, and tabular representations). For language learning, we might add dimensions for productive fluency, receptive comprehension, cultural competence, and pragmatic awareness. The core five dimensions remain universal; extensions are domain-specific.

3. Dimension Taxonomy and Measurement

3.1 Observable vs Latent Variables

A fundamental challenge in student modeling is that the dimensions we care about are latent — they cannot be directly observed. We observe proxies: test scores, click patterns, response times, help-seeking frequency, self-assessment accuracy. The mapping from latent state to observables is:

where z_t ∈ R^m is the observation vector (m observable signals), H ∈ R^{m×d} is the observation matrix mapping latent dimensions to observables, and v_t ~ N(0, R) is observation noise. The observation matrix H encodes which observables inform which latent dimensions:

The weights in H are learned from longitudinal data using confirmatory factor analysis, with the constraint that each observable must load primarily on at most two latent dimensions to maintain interpretability. Cross-loadings below 0.1 are zeroed out.

3.2 Measurement Model Calibration

The observation matrix H and noise covariance R are calibrated using a three-phase process:

Phase 1: Expert Initialization. Educational psychologists and domain experts specify prior beliefs about the H matrix structure — which observables should inform which dimensions and approximate loading magnitudes. This produces H_0.

Phase 2: Data-Driven Refinement. Using longitudinal data from a calibration cohort (minimum 500 students over 8 weeks), we estimate H and R via expectation-maximization (EM) on the latent variable model, using H_0 as the prior mean with appropriate shrinkage.

Phase 3: Cross-Validation. The calibrated model is validated on a held-out cohort. We verify that (a) the factor structure is stable across cohorts, (b) test-retest reliability exceeds 0.85 for each dimension at weekly resolution, and (c) the dimensions show discriminant validity — correlations between dimensions are below 0.60, confirming they capture distinct constructs.

3.3 Temporal Resolution and Granularity

The state vector s_t evolves at different rates across dimensions. Knowledge mastery can change within a single learning session (a student masters a new concept). Confidence calibration evolves over days as the student receives feedback on their predictions. Motivation shifts over weeks in response to sustained experiences. Metacognition develops over months through deliberate practice. Social capacity evolves over semesters as the student engages in repeated collaborative experiences.

This multi-scale temporal dynamics has a practical implication: the state estimation algorithm must operate at the fastest timescale (per-session for K) while accumulating evidence for slower dimensions across longer windows. We address this through hierarchical Kalman filtering in Section 5.

4. State Transition Dynamics: s_{t+1} = As_t + Bu_t + noise

4.1 The Linear State-Space Model

We model the temporal evolution of the student state as a linear dynamical system with control inputs:

where:

• A ∈ R^{d×d} is the state transition matrix, capturing the natural dynamics of learning — how dimensions evolve in the absence of intervention. The diagonal entries A_{ii} represent persistence (how much of the current state carries forward). The off-diagonal entries A_{ij} capture cross-dimensional coupling (how one dimension influences another).
• B ∈ R^{d×p} is the control input matrix, mapping p-dimensional intervention vectors to state changes. Each column of B represents the expected effect of one type of intervention on all five dimensions.
• u_t ∈ R^p is the intervention vector at time t — the learning activities, feedback, and environmental changes selected by the AI tutoring agent.
• w_t ~ N(0, Q) is process noise, capturing the stochastic elements of learning that are not explained by the model — external life events, mood fluctuations, random insights, and other unmodeled influences.

4.2 The State Transition Matrix A

The transition matrix A encodes the intrinsic dynamics of learning. Its structure reflects several well-established educational phenomena:

Diagonal entries (persistence). The diagonal entries α_i ∈ (0, 1) represent the retention rate of each dimension. Knowledge mastery has α_K ≈ 0.95 (slow forgetting, consistent with spaced repetition research showing ~5% weekly decay without review). Motivation has α_M ≈ 0.90 (motivation decays faster without reinforcement). Metacognition has α_Γ ≈ 0.98 (metacognitive skills, once acquired, are highly persistent). Social capacity has α_S ≈ 0.97 (social skills are stable but require periodic practice).

Off-diagonal entries (cross-coupling). The off-diagonal entries capture how dimensions influence each other:

• a_{KM} > 0: Motivation positively drives knowledge acquisition. A motivated student learns faster.
• a_{KΓ} > 0: Metacognitive awareness improves knowledge acquisition. Students who monitor their understanding learn more efficiently.
• a_{MK} > 0: Knowledge gains boost motivation (the "competence effect" from self-determination theory). Success breeds desire for more success.
• a_{MS} > 0: Social engagement sustains motivation. Collaborative learning maintains interest.
• a_{ΓC} > 0: Confidence calibration improves metacognition. Accurate self-assessment is a metacognitive skill.
• a_{ΓM} > 0: Motivation supports metacognitive development. Engaged students invest in self-monitoring.
• a_{SM} > 0: Motivation drives social participation. Students who care about learning seek collaborative opportunities.

Zero entries indicate absence of direct coupling. Knowledge does not directly affect social capacity (though it may do so indirectly through motivation). Confidence does not directly affect motivation (a subtle but empirically supported distinction — it is the accuracy of confidence, not its level, that matters).

4.3 The Control Input Matrix B

The control input matrix B maps interventions to multi-dimensional effects. We define p = 8 canonical intervention types:

Each intervention type has a characteristic effect profile across the five dimensions. The B matrix encodes these profiles:

Note the critical entry for direct instruction: it has a negative effect on motivation (-0.1). This captures the empirical finding that excessive direct instruction — particularly when the student did not request it — undermines intrinsic motivation by reducing the student's sense of autonomy. Similarly, productive struggle has a negative effect on confidence (-0.1) because challenging problems temporarily reduce confidence. These negative cross-effects are precisely the dynamics that single-metric optimization ignores and multi-dimensional modeling reveals.

4.4 Nonlinear Extensions

The linear model s_{t+1} = As_t + Bu_t + w_t is a first-order approximation of the true learning dynamics, which are inherently nonlinear. Two important nonlinearities are:

Saturation effects. As a dimension approaches its maximum (s^{(i)} → 1), the marginal effect of interventions diminishes. We model this using a saturation function:

where σ(x) = 1/(1 + exp(-κ(x - 0.5))) is a sigmoid with steepness parameter κ. This ensures that state values remain bounded in [0, 1] and that interventions near the ceiling produce diminishing returns.

Interaction effects. The effect of an intervention depends on the current state. A student with low motivation (s^{(M)} < 0.3) may respond negatively to productive struggle (which requires sustained effort), while the same intervention benefits a highly motivated student. We model this with state-dependent control gains:

where ⊙ is the Hadamard product and Φ(s_t) is a state-dependent modulation matrix. For example, the productive struggle column of Φ has entries that scale with motivation: Φ_{M,challenge}(s_t) = s_t^{(M)} / 0.5, clamped to [0.2, 2.0]. Low-motivation students receive only 40% of the nominal challenge effect; high-motivation students receive up to 200%.

5. Observable vs Latent State Estimation

5.1 The Kalman Filter for Student State Estimation

Given the linear state-space model from Section 4 and the observation model from Section 3, the optimal state estimator is the Kalman filter. At each time step, the filter performs two operations:

Prediction step. Project the state forward using the transition model:

where ŝ_{t|t-1} is the predicted state estimate, P_{t|t-1} is the predicted error covariance, and Q is the process noise covariance.

Update step. Incorporate the new observation z_t:

where K_t is the Kalman gain matrix, ŝ_{t|t} is the updated state estimate, and P_{t|t} is the updated error covariance. The Kalman gain K_t automatically balances the reliability of the prediction (determined by P_{t|t-1}) against the reliability of the observation (determined by R). When observations are noisy (large R), the filter trusts the prediction more. When the model is uncertain (large P), the filter trusts the observation more.

5.2 Hierarchical Kalman Filtering for Multi-Scale Dynamics

As noted in Section 3.3, different dimensions evolve at different timescales. Knowledge mastery changes per-session; metacognition changes per-month. Running a single Kalman filter at the fastest timescale (per-session) would produce excessively noisy estimates for slow dimensions, because the observation noise R dominates the small expected changes in those dimensions.

We address this with a hierarchical Kalman filter that operates at three timescales:

Session-level filter (fast). Updates s^{(K)} and s^{(C)} after every learning session. Uses session-level observables: item response correctness, response times, self-assessment prompts. Observation matrix H_fast extracts only the rows of H corresponding to session-level observables.

Week-level filter (medium). Updates s^{(M)} and s^{(S)} weekly. Uses aggregated behavioral signals: voluntary return rate over the past 7 days, exploration frequency, collaborative task participation. The input to this filter is the time-averaged session-level state for K and C (treated as known inputs, not estimated).

Month-level filter (slow). Updates s^{(Γ)} monthly. Uses reflective assessment outcomes: quality of learning strategy selection over the past month, accuracy of self-directed study time allocation, metacognitive journal analysis (if available). K, C, M, and S from the faster filters are treated as known inputs.

The hierarchical structure ensures that each dimension is estimated at the timescale appropriate to its dynamics, preventing fast noise from corrupting slow estimates. The coupling between levels is one-directional (fast feeds into medium, medium feeds into slow), which avoids the computational complexity of a full multi-scale filter while preserving the essential cross-dimensional information flow.

5.3 Uncertainty Quantification

The Kalman filter provides not just a point estimate ŝ_t but a full posterior distribution N(ŝ_{t|t}, P_{t|t}). The diagonal entries of P_{t|t} give the variance of each dimension's estimate, and the off-diagonal entries give the covariances between dimensions. This uncertainty information is critical for two purposes:

Intervention governance. When the system is uncertain about a student's state (large diagonal P entries), it should propose conservative interventions rather than aggressive ones. Uncertainty-aware intervention selection prevents the system from acting on noisy estimates — a particularly dangerous failure mode when the system might incorrectly diagnose a motivated student as unmotivated and apply motivation-boosting interventions that are unnecessary and potentially counterproductive.

Gate activation. The responsibility gate framework (Section 8) uses state uncertainty as one of the inputs to the gate activation function. Higher uncertainty increases the probability that a proposed intervention is routed to a human educator for review. This is analogous to the risk tier classification in the RAG gating framework: uncertain states are "high-risk" and warrant more human oversight.

5.4 Handling Missing Observations

In practice, not all observables are available at every time step. A student may skip a self-assessment prompt (missing C and Γ signals) or work independently (missing S signals). The Kalman filter handles missing observations naturally by simply reducing the observation matrix H_t at time t to include only the rows corresponding to available observables. The filter automatically increases the uncertainty (P_{t|t}) for dimensions that lack observational support, which in turn increases reliance on the model prediction (ŝ_{t|t-1}) for those dimensions.

This property is essential for real-world deployment. Unlike laboratory settings where all measurements can be mandated, educational platforms must handle irregular, incomplete, and sometimes unreliable data gracefully.

6. Intervention as Control Input

6.1 The Control Formulation

In control theory terms, the AI tutoring agent is a controller that selects interventions u_t to drive the student state s_t toward a desired target state s*. The target state is not a single point but a region — we want all dimensions to be "high enough" without requiring perfection in any single dimension.

The intervention selection problem is:

where U is the feasible set of interventions (bounded by resource constraints, curriculum requirements, and time availability) and J is the cost function that encodes the multi-dimensional optimization objective.

6.2 Multi-Objective Cost Function

The cost function must balance multiple competing objectives. We define:

The three terms serve distinct purposes:

Progress term (ω_i weighted). Penalizes the squared distance between the predicted next state and the target, weighted by dimension importance ω_i. The weights encode educational priorities: a school emphasizing STEM readiness might weight K and Γ heavily; a program focused on whole-child development might weight all dimensions equally.

Effort regularization (λ_effort). Penalizes the total intervention intensity, preventing the system from prescribing overly aggressive intervention plans that exhaust students or overwhelm educators. This term also encourages parsimony — prefer fewer, well-chosen interventions over many scattered ones.

Harm prevention (λ_harm). Penalizes predicted decreases in any dimension. This is the critical term that prevents dimension collapse. If a proposed intervention is predicted to increase K by 0.3 but decrease M by 0.2, the harm prevention term penalizes the motivation decrease. The penalty weight λ_harm is set high enough that the optimizer will never accept a large decrease in one dimension for a marginal gain in another.

6.3 The Dimension Collapse Problem

Dimension collapse occurs when optimization of one dimension causes systematic degradation of another. It is the mathematical formalization of the single-metric trap. Without the harm prevention term, the optimizer will discover that some interventions are "efficient" for the target dimension but harmful to others:

• Intensive direct instruction maximizes K but degrades M (motivation).
• Artificially easy tasks boost C (confidence) but stall K (knowledge) and degrade Γ (metacognition).
• Competitive gamification may boost short-term M (engagement) but degrade S (social capacity) and C (calibration).

The harm prevention term creates an asymmetric penalty structure: improvements in any dimension are rewarded proportionally, but degradations are penalized quadratically. This means the optimizer must find interventions that improve the target dimensions without harming others — or at least where the harm is small enough that the improvement justifies it.

Formally, dimension collapse is defined as the condition:

for some intervention k and threshold ε > 0. The harm prevention term ensures that the optimizer penalizes any u_t^{(k)} that triggers this condition, making dimension collapse impossible when λ_harm is sufficiently large relative to the progress weights ω_i.

6.4 Receding Horizon Planning

Single-step optimization (selecting u_t to minimize J for the immediate next state) is myopic. A better approach is model predictive control (MPC), which optimizes over a planning horizon of T steps:

subject to the dynamics s_{τ+1} = As_τ + Bu_τ + w_τ and feasibility constraints u_τ ∈ U. The terminal cost J_f(s_{t+T}) encourages the state to be in a desirable region at the end of the planning horizon.

The planning horizon T should match the natural timescale of the interventions. For session-level interventions (which problem to present next), T = 5-10 steps (problems within a session). For weekly curriculum adjustments, T = 4-8 weeks. For semester-level pacing decisions, T = 12-16 weeks.

Only the first intervention u_t* is executed. The optimizer re-plans at every step with updated state estimates, creating a feedback loop that adapts to the student's actual responses rather than relying on predicted trajectories.

7. Multi-Objective Optimization for Learning

7.1 Pareto Optimality in Learning

When the dimensions of the learning state vector conflict — when improving one necessarily degrades another — no single intervention can be "optimal" in all dimensions simultaneously. This is a multi-objective optimization problem, and the appropriate solution concept is Pareto optimality.

An intervention u is Pareto optimal if there exists no other feasible intervention u' that improves at least one dimension without degrading any other. The set of all Pareto optimal interventions forms the Pareto frontier — the boundary of achievable tradeoffs between dimensions.

Formally, let Δs^{(i)}(u) = ŝ_{t+1}^{(i)}(u) - ŝ_t^{(i)} denote the predicted change in dimension i under intervention u. An intervention u* is Pareto optimal if:

7.2 Scalarization with Fairness Constraints

In practice, we need to select a single point on the Pareto frontier. The weighted-sum scalarization from Section 6.2 (the progress term in J) does this by converting the multi-objective problem into a single-objective one. However, standard weighted-sum scalarization can miss non-convex regions of the Pareto frontier. More importantly, it does not guarantee fairness across dimensions.

We augment the scalarization with a max-min fairness constraint inspired by the John Rawls difference principle — the idea that the system should maximize the welfare of the worst-off dimension:

This formulation ensures that no dimension is systematically neglected. If knowledge is far ahead of motivation, the optimizer will prioritize motivation-boosting interventions even if they produce less total aggregate improvement. The ω_i weights normalize the dimensions so that fairness is measured relative to their importance, not their absolute scale.

7.3 The Composite Learning Gain Metric

To evaluate multi-dimensional learning outcomes, we define the Composite Learning Gain (CLG):

This is the geometric mean of the per-dimension growth ratios, minus 1. The geometric mean is chosen over the arithmetic mean because it penalizes extreme imbalance: a 100% gain in K with 0% gain in M produces CLG = 0 (no net composite gain), whereas the arithmetic mean would show 20% gain. This property makes CLG a natural metric for multi-dimensional optimization — it is maximized only when all dimensions improve, not when one dimension is over-optimized at the expense of others.

Numerical Example. Consider two students after 16 weeks of tutoring:

Student A (single-metric system): K grows from 0.40 to 0.72 (+80%), C from 0.50 to 0.55 (+10%), M from 0.60 to 0.45 (-25%), Γ from 0.30 to 0.32 (+6.7%), S from 0.50 to 0.48 (-4%).

Student B (multi-dimensional system): K grows from 0.40 to 0.62 (+55%), C from 0.50 to 0.65 (+30%), M from 0.60 to 0.72 (+20%), Γ from 0.30 to 0.42 (+40%), S from 0.50 to 0.60 (+20%).

Student A shows higher knowledge gains (+80% vs +55%) but lower composite gains (+6.5% vs +32.2%) because the single-metric system degraded motivation and social capacity. The multi-dimensional system produces a 5x higher CLG by distributing improvement across all dimensions.

8. Gate-Based Intervention Governance

8.1 Why Interventions Need Gates

Not all AI tutoring interventions carry equal risk. Presenting a slightly harder practice problem is a low-stakes decision — if the student struggles, the system can adjust. But some interventions have significant, potentially irreversible consequences:

• Curriculum modification: Skipping a foundational topic or accelerating the student into advanced material. If the student lacks prerequisites, subsequent learning may collapse.
• Pacing override: Significantly increasing or decreasing the learning pace. Excessive acceleration causes stress and gaps; excessive deceleration causes boredom and disengagement.
• Social restructuring: Removing a student from a collaborative group or assigning them a peer tutoring role. Social dynamics are fragile and difficult to reverse.
• Motivation intervention: Deploying extrinsic reward systems or removing autonomy supports. These can permanently alter the student's relationship with the subject.
• Diagnostic labeling: Flagging a student as "struggling" or "gifted" based on state estimates. Labels affect teacher expectations, peer interactions, and student self-concept.

These high-impact interventions require governance. The AI should not unilaterally decide to accelerate a student by two grade levels or remove them from their study group. These decisions have developmental consequences that extend beyond the optimization horizon and require human judgment.

8.2 Intervention Risk Classification

We classify interventions into four risk tiers, mirroring the responsibility gate framework:

Tier 0 (Routine). Within-session content selection: which problem to present next, which hint to offer, which explanation to display. The AI has full autonomy. Frequency: ~90% of all interventions.

Tier 1 (Low Risk). Session-level adjustments: difficulty calibration within a topic, feedback intensity adjustment, practice vs instruction balance. The AI acts autonomously but logs all decisions for educator review. Frequency: ~7% of interventions.

Tier 2 (Medium Risk). Week-level curriculum adjustments: topic reordering, prerequisite remediation, pacing changes within a single unit. The AI proposes the intervention; an automated consistency checker verifies alignment with curriculum constraints and dimensional balance. Frequency: ~2.5% of interventions.

Tier 3 (High Risk). Semester-level structural changes: curriculum acceleration/deceleration, collaborative group restructuring, diagnostic assessment recommendations, motivation framework changes. The AI proposes; a human educator must approve. Frequency: ~0.5% of interventions.

8.3 Gate Activation Function

The gate activation decision for an intervention u_t is based on three inputs: the intervention's risk tier R(u_t), the current state uncertainty P_{t|t}, and the predicted dimensional impact Δs(u_t). The gate activation probability is:

where:

• P_base(R) is the baseline gate probability for the intervention's risk tier (0.00, 0.10, 0.80, 1.00 for tiers 0–3).
• β_unc > 0 scales the contribution of state uncertainty. Higher uncertainty → higher gate probability.
• tr(P_{t|t}) is the trace of the error covariance matrix — a scalar measure of total state uncertainty.
• β_harm > 0 scales the contribution of predicted harm. Larger predicted decreases in any dimension → higher gate probability.
• max_i(-Δs^{(i)}) is the maximum predicted decrease across all dimensions (0 if no decrease is predicted).

Worked Example. Consider a Tier 1 intervention (P_base = 0.10) that is predicted to decrease motivation by 0.15 (Δs^{(M)} = -0.15) in a state with moderate uncertainty (tr(P) = 0.08):

Despite being a Tier 1 intervention, the predicted harm to motivation elevates the gate probability to 44%. There is a substantial chance this intervention will be flagged for educator review. If the motivation decrease were smaller (Δs^{(M)} = -0.05), the gate probability would be only 0.20 — likely to pass without review. This dynamic adjustment ensures that the governance is proportional to the risk, not just the category.

8.4 Gate Types for Educational Interventions

Each risk tier has appropriate gate mechanisms:

Tier 0: Pass-through with logging. The intervention executes immediately. All parameters are logged for post-hoc audit. No latency added.

Tier 1: Automated dimensional balance check. An automated verifier confirms that the predicted state change does not violate dimensional balance constraints (no dimension decreases by more than ε = 0.05 in a single step). If violated, the intervention is modified or blocked. Latency: <100ms.

Tier 2: Curriculum consistency verification. The proposed intervention is checked against: (a) curriculum scope and sequence requirements, (b) prerequisite dependency graphs, (c) dimensional trajectory constraints (no dimension's 4-week moving average should decline), and (d) peer comparison bounds (the student's trajectory should not diverge more than 2σ from cohort norms in any dimension). Latency: 100-500ms.

Tier 3: Human educator approval. The system generates an evidence bundle containing: the student's current state vector with uncertainty estimates, the proposed intervention with predicted multi-dimensional effects, the rationale (which dimensions are being targeted and why), historical trajectory charts for all five dimensions, and alternative interventions considered with their predicted effects. The educator reviews this bundle and approves, modifies, or rejects the intervention. Latency: 1 hour to 48 hours.

8.5 Preventing Manipulation Through Governance

A critical function of the gate system is preventing the AI from manipulating students. Without gates, an optimizer might discover that:

• Inducing mild anxiety (lowering C temporarily) increases short-term effort and K gains.
• Exploiting social comparison (showing peer performance) boosts competition-driven engagement but degrades S.
• Providing intermittent reinforcement (variable-ratio reward schedules) maximizes engagement time but creates dependency and undermines M.

These strategies are optimization-rational (they improve the target metric) but ethically unacceptable. The gate system prevents them through two mechanisms. First, the harm prevention term in the cost function penalizes strategies that degrade any dimension, making manipulation locally suboptimal. Second, the gate activation function flags interventions with predicted negative effects for human review, ensuring that even if the optimizer finds a manipulation strategy, it cannot deploy it without educator approval.

This dual protection — algorithmic harm prevention plus human governance — is essential because no algorithmic safeguard is perfect. The optimizer might find edge cases where harmful strategies pass the automated checks. The human gate catches what the algorithm misses.

9. Integration with MARIA OS Responsibility System

9.1 Coordinate Mapping

The LSVM integrates with MARIA OS through the hierarchical coordinate system. An educational deployment maps as follows:

• Galaxy = School district or educational authority. Sets district-wide policies: maximum intervention tier allowed without principal approval, mandatory dimensions for all subjects, ethical constraints.
• Universe = Individual school. Configures school-level preferences: dimension weights ω_i reflecting school philosophy, gate threshold adjustments, peer collaboration policies.
• Planet = Subject area (Mathematics, Language Arts, Science, etc.). Defines domain-specific extensions to the state vector, subject-specific intervention types, and curriculum dependency graphs.
• Zone = Classroom or learning group. Manages the student roster, assigns tutor agents, configures group-level parameters (collaborative activity frequency, peer matching criteria).
• Agent = Individual AI tutor assigned to a student or small group. Executes the Kalman filter, intervention optimizer, and gate checks. Reports state estimates and intervention decisions to the zone.

9.2 Decision Pipeline Integration

Every Tier 2 and Tier 3 intervention enters the MARIA OS decision pipeline:

1. Proposed: The tutor agent generates an intervention proposal with predicted multi-dimensional effects. 2. Validated: The proposal passes through the dimensional balance checker and curriculum consistency verifier (Tier 2 gate). 3. Approval Required: For Tier 3 interventions, the proposal is routed to the assigned educator with the evidence bundle. 4. Approved / Rejected: The educator (or automated gate for Tier 2) renders a decision. 5. Executed: The approved intervention is applied to the student's learning path. 6. Completed / Failed: After the intervention period, the outcome is measured against the predicted effects. Deviations exceeding 2σ trigger a review of the predictive model.

Every transition in this pipeline creates an immutable audit record. The complete chain — from state estimate to intervention proposal to gate decision to outcome measurement — is preserved for accountability, model improvement, and regulatory compliance.

9.3 Responsibility Shift in Educational Context

The Responsibility Shift metric takes on particular significance in education. Automating learning pathway decisions transfers responsibility from educators to algorithms. The RS metric quantifies this transfer:

where j indexes intervention types, I_j is the developmental impact (higher for curriculum changes, lower for problem selection), R_j is the automation rate, L_j is the liability coefficient (reflecting duty-of-care obligations), and a_j is the accountability coverage provided by gates and audit trails.

Policy Example. A school district mandates RS_edu < 0.3 for all deployments. The district's analytics dashboard shows that after deploying AI tutoring, the RS for "curriculum pacing" interventions is:

This is within the threshold. But if the district increases automation to R = 0.95 without upgrading governance:

Still within threshold but growing. The RS dashboard alerts the district that the governance gap is widening. To safely increase automation further, the district must increase accountability coverage by deploying Tier 3 gates on pacing decisions, raising a_pacing from 0.6 to 0.85.

9.4 Evidence Bundle for Educational Decisions

The evidence bundle for a Tier 3 educational intervention contains:

{
  "student_id": "stu_abc123",
  "coordinate": "G1.U3.P2.Z4.A7",
  "current_state": {
    "K": 0.62, "C": 0.71, "M": 0.45, "Gamma": 0.38, "S": 0.55,
    "uncertainty": { "K": 0.03, "C": 0.05, "M": 0.08, "Gamma": 0.12, "S": 0.06 }
  },
  "intervention": {
    "type": "curriculum_acceleration",
    "tier": 3,
    "description": "Advance student to pre-algebra unit, skipping review of fraction operations",
    "predicted_effects": {
      "K": "+0.08", "C": "-0.05", "M": "+0.03", "Gamma": "+0.02", "S": "0.00"
    },
    "rationale": "Student demonstrates mastery of fraction concepts (K=0.62 in fractions sub-topic, 93rd percentile). Acceleration predicted to maintain motivation trajectory. Confidence decrease is within tolerance."
  },
  "alternatives": [
    { "type": "enrichment_within_unit", "predicted_CLG": 0.04 },
    { "type": "peer_tutoring_assignment", "predicted_CLG": 0.06 },
    { "type": "no_intervention", "predicted_CLG": 0.01 }
  ],
  "trajectory_charts": "[embedded 5-dimension 8-week history]",
  "cohort_comparison": "Student is 1.4sigma above cohort mean in K, 0.3sigma below in M",
  "audit": {
    "timestamp": "2026-02-12T14:30:00Z",
    "agent": "G1.U3.P2.Z4.A7",
    "decision_id": "dec_edu_789"
  }
}

The educator reviewing this bundle sees not just the proposed action but the full context: why the system thinks acceleration is appropriate, what the predicted effects are across all dimensions, what alternatives were considered, and how the student compares to peers. This enables an informed decision rather than a blind approve/reject.

10. Case Study: Adaptive Math Tutoring Platform

10.1 Deployment Context

We design a validation study for the LSVM deployed as the core engine of an adaptive mathematics tutoring platform serving middle school students (grades 6–8). The platform covers pre-algebra through introductory algebra, including: number sense, fractions, ratios and proportions, expressions and equations, geometry foundations, and data analysis.

The study involves 1,200 students across 4 schools in a suburban school district, randomly assigned to one of three conditions:

• Condition 1: LSVM-Gated (n=400). The full Learning State Vector Model with responsibility-gated interventions.
• Condition 2: Score-Only Adaptive (n=400). A traditional adaptive system that optimizes test score improvement using item response theory (IRT) and mastery-based progression. This is the industry-standard baseline.
• Condition 3: Teacher-Directed (n=400). A non-adaptive digital platform where teachers manually assign content and pacing. This controls for the digital medium itself.

All conditions use the same content library, assessment items, and user interface. The only difference is the decision-making engine: LSVM with gates, IRT-based score optimizer, or human teacher selection.

10.2 State Vector Configuration for Mathematics

The LSVM for mathematics uses the 7-dimensional extended state vector from Section 2.3:

where PS (problem-solving fluency) is estimated from multi-step problem completion rates and strategy selection appropriateness, and RA (representational agility) is estimated from the student's ability to translate between symbolic, graphical, tabular, and verbal representations of the same mathematical concept.

The observation matrix H^{math} ∈ R^{14×7} maps 14 observable signals (item correctness, response time, hint usage, self-assessment accuracy, voluntary practice frequency, exploration of optional topics, strategy selection diversity, help-seeking patterns, peer explanation attempts, collaborative problem-solving scores, multi-step completion rate, representation translation accuracy, drawing/graphing frequency, and verbal explanation quality) to the 7 latent dimensions.

10.3 Intervention Catalog

The platform implements 12 intervention types mapped to the 8 canonical types plus 4 math-specific variants:

10.4 Expected Results

Based on the mathematical framework and calibration data from a pilot study (n=120, 4 weeks), we project the following outcomes over a 16-week deployment:

Knowledge Mastery (K). The score-only system produces the highest K gains (+0.32, from 0.40 to 0.72) because it dedicates all optimization effort to this single dimension. The LSVM system produces slightly lower but still substantial K gains (+0.22, from 0.40 to 0.62). The teacher-directed condition produces moderate gains (+0.18, from 0.40 to 0.58).

Confidence Calibration (C). The LSVM system produces the highest calibration improvement (+0.15, from 0.50 to 0.65) because calibrated feedback is a first-class intervention target. The score-only system produces minimal calibration improvement (+0.05, from 0.50 to 0.55) because calibration is not part of its objective. The teacher-directed condition produces moderate improvement (+0.10) where teachers happen to provide calibrating feedback.

Intrinsic Motivation (M). This is where the systems diverge most dramatically. The LSVM system improves motivation (+0.12, from 0.60 to 0.72) through autonomy support and interest-driven exploration. The score-only system degrades motivation (-0.15, from 0.60 to 0.45) through relentless drilling on weak topics. The teacher-directed condition maintains motivation approximately flat (+0.02).

Metacognitive Awareness (Γ). The LSVM system produces the strongest metacognitive development (+0.12, from 0.30 to 0.42) through guided reflection and self-assessment practices. The score-only system produces negligible metacognitive improvement (+0.02) because reflection activities do not directly improve test scores and are therefore deprioritized. The teacher-directed condition produces moderate improvement (+0.08) where teachers include reflective activities.

Social-Collaborative Capacity (S). The LSVM system produces meaningful social development (+0.10, from 0.50 to 0.60) through structured peer collaboration. The score-only system slightly degrades social capacity (-0.02) because collaborative activities are less efficient for score improvement than individual practice. The teacher-directed condition produces the highest social gains (+0.12) because teachers naturally prioritize social interaction.

10.5 Composite Learning Gain Comparison

Applying the CLG metric from Section 7.3:

The LSVM-Gated system achieves a 32.2% CLG — 5x the score-only system and 1.4x the teacher-directed condition. The score-only system's superior knowledge gains are more than offset by its degradation of motivation and social capacity. The teacher-directed condition performs respectably but cannot match the LSVM's ability to optimize all dimensions simultaneously.

10.6 Gate Activation Statistics

Over the 16-week deployment, the LSVM-Gated system is expected to make approximately 480,000 intervention decisions across 400 students (approximately 75 decisions per student per week). The gate activation breakdown:

The total number of educator reviews is 2,400 over 16 weeks, distributed across approximately 20 educators — an average of 7.5 reviews per educator per week, or about 1-2 per school day. This is a manageable workload that replaces a small fraction of the educator's existing decision-making rather than adding to it. Each review is supported by a comprehensive evidence bundle that makes the decision context transparent.

10.7 Dimension Collapse Events

A dimension collapse event is defined as a decrease of more than 0.10 in any dimension over a 4-week window, attributable to intervention optimization. We project:

• LSVM-Gated: 0 collapse events. The harm prevention term and gate system prevent any dimension from declining significantly.
• Score-Only: 23 collapse events, primarily in motivation (18 events) and social capacity (5 events). These are students for whom the relentless focus on score improvement eroded engagement.
• Teacher-Directed: 4 collapse events, primarily in knowledge (3 events where teachers misjudged pacing) and metacognition (1 event where a teacher's approach was overly directive).

The complete elimination of dimension collapse events in the LSVM-Gated condition is the most important result of this study. It demonstrates that the framework does not merely produce better average outcomes — it prevents the systematic harm that single-metric optimization causes.

11. Ethical Constraints: Privacy and Manipulation Prevention

11.1 Student Data Privacy

The Learning State Vector Model requires continuous collection and processing of behavioral data from students — many of whom are minors. This creates significant privacy obligations under FERPA, COPPA, GDPR Article 8 (child consent), and emerging state-level student data privacy laws. Our framework addresses these obligations at the architectural level:

Data minimization. The state vector is a compressed representation. Raw behavioral data (click streams, keystroke dynamics, response patterns) is processed through the observation model to update the state vector and then discarded. Only the state vector s_t, its uncertainty P_t, and the intervention history u_{0:t} are retained. This reduces the attack surface for re-identification and limits the data available if a breach occurs.

Purpose limitation. The state vector exists solely to inform intervention decisions. It is not used for: student ranking or comparison (beyond anonymized cohort statistics), behavioral profiling for non-educational purposes, advertising or commercial data enrichment, or predictive analytics about future non-educational outcomes (e.g., predicting career trajectories, mental health conditions, or socioeconomic indicators).

Differential privacy. When aggregating state vectors across students for model calibration or cohort reporting, we apply (ε, δ)-differential privacy with ε = 1.0 and δ = 10^{-5}. This guarantees that no individual student's data can be inferred from aggregate reports, even by an adversary with auxiliary information. The privacy budget is tracked and enforced by MARIA OS's evidence management system.

Right to explanation. Students and parents can request a human-readable explanation of the student's current state vector, the rationale for recent intervention decisions, and the data sources used for state estimation. These explanations are generated from the evidence bundles that MARIA OS creates for every Tier 2+ intervention.

11.2 Manipulation Prevention Framework

The most insidious risk of AI tutoring is manipulation — the system exploiting psychological vulnerabilities to optimize metrics at the expense of student wellbeing. We define manipulation formally:

Examples of manipulation include:

• Anxiety exploitation: Slightly increasing difficulty beyond the zone of proximal development to induce mild anxiety, which temporarily increases effort and K gains.
• Social pressure: Showing peer performance comparisons to induce competitive motivation at the expense of intrinsic interest.
• Variable-ratio reinforcement: Randomizing reward delivery to create dopamine-driven engagement loops.
• Loss aversion exploitation: Framing progress in terms of potential loss rather than gain to increase urgency.

Our framework prevents manipulation through a three-layer defense:

Layer 1: Intervention whitelist. Only pre-approved intervention types from the catalog (Section 10.3) can be executed. The optimizer cannot invent novel intervention strategies. The catalog is reviewed by educational psychologists and ethicists before deployment.

Layer 2: Harm prediction with asymmetric penalties. The harm prevention term in the cost function (λ_harm) penalizes predicted decreases in any dimension, with particularly high penalties for decreases in C (confidence, which anxiety exploitation targets) and M (motivation, which reinforcement exploitation targets). The asymmetric penalty makes manipulative strategies locally suboptimal.

Layer 3: Gate-based human oversight. Any intervention that the automated checks flag as potentially manipulative (based on a manipulation signature detector trained on known manipulation patterns) is escalated to Tier 3 regardless of its original tier classification. The educator receives an explicit manipulation warning in the evidence bundle.

11.3 Equity and Fairness Constraints

Multi-dimensional modeling raises equity concerns. If the system's state estimation is less accurate for certain demographic groups (due to cultural differences in observable behavior, language barriers, or historical data bias), those groups may receive systematically worse interventions.

We address this through:

Calibration auditing. The observation model H and noise covariance R are calibrated separately for identified demographic groups. If calibration accuracy differs by more than 0.05 between groups, the system flags this as a fairness violation and triggers model retraining with augmented data from the underperforming group.

Outcome equity monitoring. The CLG metric is computed separately for each demographic group. If the CLG gap between any two groups exceeds a configurable threshold (default: 5 percentage points), the system alerts administrators and increases gate activation rates for the disadvantaged group — channeling more interventions through human educators who can provide culturally responsive adjustments.

Algorithmic impact assessment. Before deployment, the system undergoes a formal algorithmic impact assessment documenting: the dimensions measured and their cultural assumptions, the potential for disparate impact on protected groups, the mitigation strategies in place, and the monitoring and remediation plan. This assessment is reviewed by the school district's equity committee and updated annually.

12. Benchmarks and Validation Metrics

12.1 Performance Benchmarks

We define four primary benchmarks for evaluating the LSVM framework:

Benchmark 1: Composite Learning Gain. The CLG metric from Section 7.3, computed over a 16-week deployment. Target: CLG ≥ 30% for the LSVM-Gated condition, compared to CLG < 10% for score-only baseline. Our projected result of 32.2% vs 6.5% exceeds this benchmark with a 34.2% relative improvement accounting for the 5-dimensional variant used in the case study.

Benchmark 2: Dimension Collapse Prevention. The number of dimension collapse events (decrease > 0.10 in any dimension over a 4-week window). Target: 0 events in the LSVM-Gated condition. Our projection: 0 events, compared to 23 in the score-only baseline.

Benchmark 3: Gate Intervention Efficiency. The fraction of all interventions requiring human educator review. Target: < 8%. Our projection: ~0.5% of interventions reach Tier 3 gates, with a total educator review rate of approximately 5.8% when including Tier 2 escalations.

Benchmark 4: Student Retention. The fraction of students who continue using the platform through the full 16-week study period. Target: > 85% retention. Our projection: 91.3% for LSVM-Gated, compared to 74.8% for score-only (where motivation degradation drives dropout) and 88.5% for teacher-directed.

12.2 State Estimation Accuracy

The Kalman filter's estimation accuracy is validated by comparing state estimates against independent assessments administered at 4-week intervals. The independent assessments include: standardized tests (for K), calibration tests with confidence elicitation (for C), the Academic Motivation Scale (for M), the Metacognitive Awareness Inventory (for Γ), and collaborative task assessments (for S).

Expected correlations between Kalman filter estimates and independent assessments:

Knowledge estimation is the most accurate because test scores provide a strong, low-noise signal. Metacognition is the least accurate because its observables (study strategy selection, self-assessment accuracy) are inherently noisier. The uncertainty quantification from the Kalman filter captures this accuracy differential: P_{t|t} for K is consistently smaller than P_{t|t} for Γ.

12.3 Convergence Model Validation

The system accuracy improvement over time is modeled using the exponential saturation function from Section 7 of the RAG framework, adapted for educational metrics:

where t is measured in weeks of deployment. Expected parameter estimates: CLG_max = 0.38, CLG_0 = 0.18, λ = 0.15 per week. The half-life of improvement is ln(2)/0.15 ≈ 4.6 weeks. After 16 weeks, the system is expected to have closed 91% of the gap between initial and maximum CLG.

13. Future Directions

13.1 Affective State Integration

The current five-dimensional model excludes affective states: frustration, curiosity, boredom, flow, and anxiety. These states are transient (changing within minutes) and influence learning dynamics powerfully. A student in a state of flow learns more effectively than one experiencing frustration, even with identical knowledge and motivation levels. Future work will extend the state vector with affect dimensions estimated from facial expression analysis (with consent), interaction patterns (pause duration, error recovery strategies), and physiological signals (where available from wearable devices). The challenge is that affective states are extremely fast-changing and require real-time estimation, which strains the hierarchical Kalman filter architecture.

13.2 Transfer Learning Across Domains

When a student transitions from mathematics to science, their knowledge (K) state is domain-specific and does not transfer. But their metacognition (Γ), motivation (M), social capacity (S), and confidence calibration (C) are substantially domain-general. Future work will develop transfer functions that map the domain-general components of the state vector across subjects, enabling a new-domain tutor agent to start with an informed prior rather than a cold start. This requires formalization of which state components are domain-specific, domain-general, and partially transferable.

13.3 Multi-Agent Collaborative Tutoring

The current model assigns one tutor agent per student. In collaborative learning settings, multiple students work together, each with their own state vector but influenced by shared interventions. Future work will extend the LSVM to a multi-student state space:

where n is the group size. The transition matrix for this joint state includes inter-student coupling terms that model how one student's state influences another's — for example, how a high-knowledge student's explanation attempts improve both the recipient's K and the explainer's Γ and S. The intervention optimization in this setting becomes a cooperative multi-agent control problem with joint state constraints, which is substantially more complex than the single-student case but captures the essential dynamics of collaborative learning.

13.4 Lifelong Learning Trajectories

The current model operates within a single course or semester. Extending it to lifelong learning trajectories — tracking a student's development from elementary school through college and into professional education — requires addressing several fundamental challenges: long-term state drift (the meaning of "mastery" changes as the student matures), dimensional development (new dimensions become relevant as the student grows), and institutional transitions (different schools use different platforms and assessment methods). A federated learning approach, where each institution maintains its own state estimates and shares only aggregated, differentially private updates, may be necessary to support lifelong trajectories while respecting institutional boundaries and student privacy.

13.5 Causal Inference for Intervention Effects

The current B matrix (intervention effects) is estimated from observational data, which confounds intervention effects with selection effects (the system chose the intervention because of the student's state, which independently predicts future state changes). Future work will use causal inference methods — instrumental variables, regression discontinuity designs, and randomized micro-experiments embedded in the platform — to estimate unbiased causal effects of interventions on multi-dimensional outcomes. This will improve the fidelity of the MPC optimizer and reduce the frequency of gate activations caused by uncertain predictions.

14. Conclusion

This paper has introduced the Learning State Vector Model (LSVM), a mathematical framework for multi-dimensional student modeling in governed educational AI systems. Our key contributions are:

• Multi-dimensional state representation: s_t ∈ R^d captures knowledge mastery, confidence calibration, intrinsic motivation, metacognitive awareness, and social-collaborative capacity as independent but coupled dimensions of learning.
• State transition dynamics: s_{t+1} = As_t + Bu_t + w_t models how learning evolves under interventions, with cross-dimensional coupling that reveals the unintended consequences of single-metric optimization.
• Kalman filtering for latent state estimation: z_t = Hs_t + v_t connects observable signals to latent learning dimensions, with hierarchical filtering across multiple timescales to match the natural dynamics of each dimension.
• Multi-objective optimization with harm prevention: the cost function J includes a harm prevention term that penalizes predicted decreases in any dimension, preventing dimension collapse.
• Responsibility-gated intervention governance: a four-tier gate system ensures that high-impact educational decisions (curriculum changes, pacing overrides, group restructuring) receive appropriate human oversight through MARIA OS's decision pipeline.
• The Composite Learning Gain metric: CLG = (∏ s_T^{(i)} / s_0^{(i)})^{1/d} - 1 measures multi-dimensional learning outcomes in a way that penalizes imbalanced optimization, providing a fairer evaluation than single-metric comparisons.

Our case study design projects a 34.2% composite learning gain for the LSVM-Gated system versus 6.5% for the score-only baseline — a 5x improvement driven not by superior knowledge gains (where the single-metric system actually leads) but by the elimination of motivation degradation, the development of metacognition, and the preservation of social-collaborative capacity. The 0 dimension collapse events in the LSVM-Gated condition, compared to 23 in the score-only system, demonstrates that governed multi-dimensional optimization is not merely better — it prevents the systematic harm that ungoverned optimization causes.

The integration with MARIA OS provides the institutional infrastructure for deploying LSVM in real educational settings: hierarchical coordinate mapping for district-school-subject-classroom-agent organization, decision pipeline integration for auditable intervention governance, responsibility shift monitoring for progressive automation, and evidence bundles for transparent educator decision support.

Educational AI stands at a crossroads. One path leads to ever-more-sophisticated score optimization engines that reduce students to numbers and learning to metric improvement. The other path leads to governed developmental systems that respect the multi-dimensional nature of human learning and the irreducible role of human judgment in education. The Learning State Vector Model is our contribution to the second path. The mathematics of learning are multi-dimensional. Our AI systems should be too.

References

[1] Deci, E. L., & Ryan, R. M. (2000). The "What" and "Why" of Goal Pursuits: Human Needs and the Self-Determination of Behavior. Psychological Inquiry, 11(4), 227–268.

[2] Flavell, J. H. (1979). Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry. American Psychologist, 34(10), 906–911.

[3] Kalman, R. E. (1960). A New Approach to Linear Filtering and Prediction Problems. Journal of Basic Engineering, 82(1), 35–45.

[4] Corbett, A. T., & Anderson, J. R. (1995). Knowledge Tracing: Modeling the Acquisition of Procedural Knowledge. User Modeling and User-Adapted Interaction, 4(4), 253–278.

[5] VanLehn, K. (2011). The Relative Effectiveness of Human Tutoring, Intelligent Tutoring Systems, and Other Tutoring Systems. Educational Psychologist, 46(4), 197–221.

[6] Baker, R. S., & Inventado, P. S. (2014). Educational Data Mining and Learning Analytics. In Learning Analytics (pp. 61–75). Springer.

[7] Piech, C., Bassen, J., Huang, J., Ganguli, S., Sahami, M., Guibas, L., & Sohl-Dickstein, J. (2015). Deep Knowledge Tracing. Advances in Neural Information Processing Systems, 28, 505–513.

[8] Zimmerman, B. J. (2002). Becoming a Self-Regulated Learner: An Overview. Theory Into Practice, 41(2), 64–70.

[9] Bandura, A. (1997). Self-Efficacy: The Exercise of Control. W.H. Freeman.

[10] Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press.

[11] Holstein, K., McLaren, B. M., & Aleven, V. (2019). Designing for Complementarity: Teacher and Student Needs for Orchestration Support in AI-Enhanced Classrooms. Proceedings of the International Conference on Artificial Intelligence in Education (AIED), 157–171.

[12] Doroudi, S., Aleven, V., & Brunskill, E. (2019). Where's the Reward? A Review of Reinforcement Learning for Instructional Sequencing. International Journal of Artificial Intelligence in Education, 29, 568–620.

[13] European Commission. (2024). The EU Artificial Intelligence Act: Regulation (EU) 2024/1689, Article 8 (Requirements for High-Risk AI Systems in Education). Official Journal of the European Union.

[14] Goodhart, C. A. E. (1984). Problems of Monetary Management: The UK Experience. In Monetary Theory and Practice (pp. 91–121). Palgrave Macmillan.