One variable repeatedly appeared near the top of the model's importance rankings: school attendance.
Students with frequent absences were substantially more likely to receive intensive support later.
The association was strong, stable, and easy to explain to officials.
That made it dangerous.
Elias had begun speaking as though poor attendance caused later academic difficulty.
Priya asked him to defend the causal claim rather than the correlation.
He immediately listed plausible mechanisms: missed instruction, weaker classroom relationships, and reduced practice.
Priya then listed mechanisms running in the opposite direction.
Existing academic difficulty could make attendance less rewarding, while illness, unsafe transport, family caregiving, or anxiety could affect both attendance and performance.
Attendance might therefore be a cause, an effect, a mediator, a symptom, or several of these at once.
The predictive model did not distinguish among those possibilities because prediction and causal inference answer different questions.
A feature can improve forecasts even when intervening on that feature would not produce the expected outcome.
To illustrate the point, Priya compared umbrellas and rain.
Umbrellas are excellent predictors of wet streets, but confiscating umbrellas would not stop rainfall.
The analogy was simple, yet its policy implication was severe.
If the city responded to high-risk predictions by pressuring students to attend without addressing the causes of absence, it might optimize the visible variable while leaving the underlying problem untouched.
Worse, punitive attendance policies could harm students whose absences reflected circumstances beyond their control.
Elias realized that model explanations are not automatically intervention recommendations.
The team began drawing causal diagrams to make assumptions explicit.
These diagrams did not prove causation, but they forced researchers to state which relationships they believed existed and which variables might confound others.
Priya insisted that any policy derived from CivicLearn distinguish predictive evidence from causal evidence.
That distinction would make official presentations less dramatic.
The causal diagrams functioned as epistemic discipline because they externalized assumptions that prose often leaves implicit and therefore difficult to challenge.
Prediction can remain empirically successful while policy intervention fails, a divergence that makes causal humility indispensable wherever forecasts influence action.
It would also make them more honest.