Watch the idea
The concept
Case
Question + allowed sources
Score
Groundedness, not style
Gate
Invented citation is caught
March is not August
New products, slang, fraud. Watch the input, then the score. Alert on the metric, not the mood of a slide.
LLM is not default
A stable classify job often wants a small model with a holdout. Fashion is not a metric.
Worked example
The quiet lie
The dashboard still said 94%. The world had moved. Drift is how production lies without a crash.
Picture to keep
Watch in · Watch out · Retrain or refuse
- 1Holdout is dated
- 2Alert on the metric
- 3LLM is not default
