Blind by Design

The Aletheia Institute is running a multi-phase vision-enhancement trial. The lab is a Spring Boot service whose one job is to record the vision_state of every subject who walks through the protocol (blurry, sharp, enhanced, or clouded), because subjects don't all arrive with the same biology, the same dose adherence, or the same trial-jurisdiction baseline. The flag definitions that drive those readings live in flags.json, watched by a flagd sidecar; the OpenFeature SDK is supposed to call that sidecar on every evaluation.

It hasn't been. For the past eight months, every subject through the door has been recorded as "untreated": the integration was never finished, and the lab director assumed the system was reading the chart. Worse, eight weeks ago the Institute opened its flagship Phase 3 trial: a new amplifier variant rolled out fractionally to a cohort by a targeting rule in flags.json. Four adverse-event reports have since been filed, each one a subject whose vision_state at discharge was worse than at enrollment.

The monitoring is dark, not by accident, but because no one ever turned the lights on. Your mission across three levels: stand up the lab so it reads the chart, read the chart by cohort so outcomes can be tracked, then turn on the lights and roll back the Phase 3 variant before the director signs off on the next enrollment batch.