Count the made-up answers.
Day 4 explained why AI invents things. This is how you measure how often it happens in your own use, so you can track it.
How to measure
Simple Method
- Take a sample of real outputs
- Check every factual claim against the source
- Count outputs with at least one invented claim
- Divide by the sample size
What to do with it
Use The Number
- Set an acceptable level per risk tier
- Track it after every change
- Compare tools on the same sample
- Lower it with grounding and clearer prompts
Measure it. Then drive it down.