CAPIRE Intervention Lab

• CAPIRE Intervention Lab is an empirically calibrated simulation environment that models curriculum constraints and attrition in higher education.
• It integrates leakage-aware data architecture, multi-level trajectory analytics, and dynamic policy experiments to assess interventions.
• Simulation results indicate that bundled policies targeting course structure and psychosocial support significantly reduce dropout rates.

The CAPIRE Intervention Lab is an empirically calibrated agent-based policy simulation environment designed to evaluate and optimize interventions in curriculum-constrained higher-education programmes, with a particular emphasis on engineering contexts characterized by high structural rigidity, persistent attrition, and exposure to macroeconomic and institutional shocks. Developed in the context of Latin American long-cycle engineering degrees, it integrates multi-level trajectory analytics, curriculum graph formalism, robust archetype discovery, and dynamic policy experimentation to address the limitations of predictive-only learning analytics frameworks (Paz, 22 Nov 2025).

1. Foundational Data Layer and Trajectory Modelling

At its core, the CAPIRE Intervention Lab is built upon a leakage-aware data architecture that enforces strict temporal alignment between predictor variables and future outcomes, eliminating data leakage and increasing the validity of both prediction and intervention simulations (Paz, 14 Nov 2025). Features are structured into four nested levels:

• N1: Personal and socio-economic attributes (e.g., deprivation indices, family educational capital).
• N2: Entry moment and academic history (e.g., secondary GPA, entry cohort, employment at entry).
• N3: Curricular friction and performance (e.g., course-level Instructional Friction Coefficient (IFC), grades, failure/dropping rates).
• N4: Institutional and macro-context variables (e.g., delays, load trends, macro-shocks).

The Value of Observation Time (VOT) formalism ensures that features used for prediction and simulation are only computed from data available up to the relevant academic term, preventing inadvertent peeking at future outcomes, and supporting causal reasoning and agent-based initialization (Paz, 14 Nov 2025).

2. Agent Initialization via Empirical Archetypes

Agents in the simulation are initialised by sampling from empirically robust trajectory archetypes, identified via a dimensionality reduction (UMAP) and density-based clustering (DBSCAN) pipeline applied to VOT-compliant, multilevel feature vectors (Paz, 22 Nov 2025, Paz, 14 Nov 2025). Thirteen archetypes were validated as stable and temporally persistent, including:

• Structurally Vulnerable: High initial failure in first-year backbone courses, accumulating blocked credits, and low sense of belonging.
• Relatively Stable: Timely gateway course passage, moderate delays, higher psychological resilience.

Each agent is parametrized by archetype membership, stress and belonging scores drawn from empirical distribution, and cohort-specific entry conditions, providing a heterogeneous and data-grounded agent population for simulation (Paz, 22 Nov 2025).

3. Curriculum Graphs and Structural Friction Indicators

The curriculum is formalised as a directed acyclic graph (DAG) $G=(V,E)$, where nodes represent courses and edges encode prerequisite constraints. Three main structural friction indicators are computed dynamically for each agent:

• Backbone Completion $BC_{i,t}$:

$BC_{i,t} = \frac{|B\cap C_{i,t}|}{|B|}$

quantifies progress through critical backbone courses.

• Blocked Credits $X_{i,t}$:

$$X_{i,t} = \sum_{v\in V,\;\mathrm{prereqs}(v)\not\subset C_{i,t}} \mathrm{credits}(v)$$

captures curriculum-imposed deadweight for agents with unsatisfied prerequisites.

• Distance to Graduation $D_{i,t}$:

$$D_{i,t} = \min_{u\in C_{i,t}} \mathrm{dist}(u,v^\ast)/\max \mathrm{dist}$$

measures shortest DAG path from completed courses to graduation.

These state variables directly modulate dropout hazard and academic evolution in the agent-based simulation, capturing the institutional topology's effect on attrition pathways (Paz, 22 Nov 2025).

4. Policy Simulation Design and Experimental Framework

The CAPIRE Intervention Lab enables systematic assessment of policy bundles through a factorial experimental design:

• A: Curriculum/Assessment Structure (status quo vs. prerequisite relaxation, modular assessment).
• B: Teaching/Academic Support (historical lectures vs. tutorials, formative feedback, remediation).
• C: Psychosocial/Financial Support (minimal vs. integrated mentoring, targeted aid).

Scenarios are evaluated in an $2\times2\times2$ factorial crossing with 100 stochastic replications per scenario, covering all combinations for a total of eight policy bundles. Each agent trajectory is simulated over twelve semesters, yielding over 12 million agent-semester records for robust statistical analysis (Paz, 22 Nov 2025).

Aggregate outcomes reveal that policy bundles targeting early backbone courses and blocked credits can reduce long-term dropout by ≈2.75 percentage points and triple the mean number of courses passed, particularly for structurally vulnerable archetypes. Teaching/support and psychosocial intervention dimensions yield the largest main effects, while curriculum redesign alone has a weaker impact unless coupled with other supports (Paz, 22 Nov 2025).

5. Macro-Shock Integration and Dual-Stressor Causal Modelling

To quantify the impact of teacher strikes (proximal shocks) and inflation (distal shocks), the CAPIRE framework incorporates a macro-shock module (Paz, 18 Nov 2025):

• Inflation indicators: annualized rate at entry, 24-month volatility, and real-time percentage changes.
• Strike exposure: proportion of class days lost, encoded at lags of 1–3 semesters and aggregated by curriculum stage.
• Dual-stressor features: interaction terms such as inflation-volatility × cumulative strike exposure.

Dropout prediction demonstrates that macro-level features raise Macro F1 from 0.73 to 0.78, with permutation tests confirming the unique contribution of inflation volatility and strike-weighted friction indices (Paz, 18 Nov 2025).

Causal inference via Double Machine Learning (LinearDML) establishes that the main effect of staff strikes is only significant when lagged by two semesters (baseline odds ratio in simple logit is 2.34), but once curriculum friction and inflation are controlled, only the interaction between lag-2 strike intensity and inflation at entry remains robust ($\hat\theta=0.0625,\ p=0.0033$), demonstrating that dropout risk from strikes is amplified nonlinearly under high inflation (Paz, 25 Nov 2025).

6. Agent-Based Macro-Shock Scenario Simulation

The CAPIRE-MACRO module embeds empirically estimated dual-shock effect sizes into the agent-based system. Agents' resilience and performance parameters are perturbed by inflation ($\alpha_{\inf}$) and strike ($\alpha_{\str}$) amplifiers. Crisis scenarios (inflation-only, strike-only, combined) are factorially crossed, and dropout under dual stressors ($D(\alpha_{\inf}, \alpha_{\str})$) is compared to additive predictions to quantify non-linear amplification: $A = D(\alpha_{\inf}, \alpha_{\str}) - \bigl[D(\alpha_{\inf}, 1) + D(1, \alpha_{\str}) - D(1, 1)\bigr]$ with combined-shock dropout rates exceeding additive expectation by 18–23% in high-intensity scenarios (Paz, 18 Nov 2025).

Scenario planning leverages these empirical and simulated results to specify actionable interventions: front-loaded financial aid for high-inflation cohorts, just-in-time academic remediation, asynchronous modules post-strike, and dynamic early-warning triggers based on predictive and causal analytics (Paz, 18 Nov 2025, Paz, 25 Nov 2025).

7. Methodological Significance and Policy Implications

The CAPIRE Intervention Lab exemplifies a shift from purely predictive, classification-based analytics to a mechanistic, theory-driven approach supporting dynamic policy design in higher education. By bridging robust feature engineering, causal inference, and agent-based simulation within a leakage-aware framework, the Lab enables institutions to address structural attrition at scale, simulate “what-if” intervention scenarios, and optimize resource allocation toward the most structurally vulnerable student archetypes (Paz, 22 Nov 2025, Paz, 14 Nov 2025, Paz, 25 Nov 2025).

This integrated pipeline also aligns with contemporary active learning and causal experimental design frameworks (Zhang et al., 2022, Gamella et al., 2020), positioning the CAPIRE Intervention Lab as a transparent, extensible sandbox for the planning and ex ante evaluation of curriculum, support, and macro-shock mitigation policies in complex institutional environments.