The redesigned CivicLearn was eventually approved for a limited deployment rather than a citywide launch.
Only a subset of schools would use it during the first semester.
The restriction disappointed officials who wanted a visible demonstration of progress.
Priya described the limitation as evidence of governance rather than hesitation.
A staged deployment allows institutions to learn from real use without exposing the entire population to an untested process.
The participating schools received training not only on the software but on uncertainty, appeals, overrides, privacy, and prohibited uses.
Students and families were informed when the system contributed to a support decision.
They were given access to an explanation, correction process, and human review.
Independent evaluators monitored outcomes using predefined metrics.
Crucially, they also tracked unintended consequences that were not part of the optimization objective.
Did students avoid counselling after learning about the system?
Did teachers alter documentation practices?
Did some schools become overly dependent on scores?
Did support resources shift away from students whose needs were less legible in data?
The evaluation combined quantitative measures with interviews and case studies.
Elias had once regarded qualitative evidence as supplementary.
He now saw that many institutional effects are invisible in aggregate statistics until someone describes how procedures actually feel and function.
After three months, the pilot showed modest improvement in identifying students who later requested support.
False-positive disparities had also narrowed.
However, one school displayed unusually high override rates.
Investigation revealed not misconduct but a local program that served recent immigrants whose records were incomplete in ways the model handled poorly.
The appropriate response was not to discipline reviewers for disagreeing with the model.
It was to recognize a population for whom the model had limited validity.
The school received a modified review process, and its data were not immediately used for retraining.
Staged deployment functioned as epistemic containment, limiting exposure while generating evidence about how the system behaved outside the assumptions of the laboratory.
The immigrant-program case illustrated external validity: a model can be sound for the population it knows and still be inappropriate for people whose histories are systematically underrepresented.
Conditional deployment had made room for a conclusion large launches often resist: sometimes the responsible output of a model is uncertainty.