The city decided not to abandon CivicLearn but to rebuild it under stricter conditions.
The redesign began by narrowing the system's purpose.
It would no longer produce a general risk score used across programs.
Instead, separate models would estimate specific forms of support need, each connected to a clearly defined intervention pathway.
This reduced the convenience of a single universal score but improved conceptual coherence.
The team replaced several proxy-heavy features with variables that were more directly related to the support being offered.
They also introduced uncertainty intervals rather than presenting every prediction as a precise point estimate.
Cases with high uncertainty would automatically receive human review.
The new interface displayed model limitations alongside recommendations.
Reviewers could see whether a student's data fell outside the range well represented in training.
Priya insisted that unfamiliar cases should trigger humility rather than extrapolation.
The team also recalibrated thresholds based on program capacity and explicit policy priorities.
Instead of hiding scarcity inside a technical cutoff, the interface showed how many students would receive support under different thresholds.
Officials therefore had to confront the resource implications of their choices.
Appeal mechanisms were integrated directly into the workflow.
Students and families could correct factual errors, provide context, request reconsideration, and receive an explanation of the outcome.
The model's score would expire after a defined period unless updated evidence justified continued use.
That prevented temporary hardship from becoming a permanent digital identity.
The team documented every major design choice in a public model card written in both technical and accessible language.
Elias worried that exposing limitations would reduce confidence in the system.
Priya argued that warranted confidence is stronger than confidence produced by omission.
The rebuilt system was less elegant than the original.
It contained more safeguards, more exceptions, more documentation, and more opportunities for disagreement.
That complexity was not failure.
The redesign substituted bounded competence for technological ambition, preferring a narrower system whose claims could be defended to a universal score whose apparent simplicity concealed conceptual overreach.
Expiration rules also rejected digital essentialism: temporary hardship would not be permitted to crystallize into a permanent computational identity.
It was the visible cost of taking the social context seriously.