Override as Signal: Learning from Human Judgment
Every time a planner overrides an algorithm, valuable information is created. This thread explores how operational overrides can become high-quality training data for future AI systems instead of disappearing once today's decision is made.
Research Question
Can a planner's overrides of an algorithm be captured, labeled, and reused as supervised signal for the next generation of the model?
Why This Matters
Most operational ML systems discard the most expensive data they generate: the moment a human says no. Treating overrides as labeled events turns every working day into a quiet training run.
Current Direction
Defining an override schema that captures context, reason, and outcome alongside the change itself, so the signal carries enough structure to be learnable later.
Early Notes
Drawing on field notes from the Procurement Heat Map Engine, where override patterns already cluster around predictable failure modes of the underlying rules.