Watch the idea
The concept
Case
Question + allowed sources
Score
Groundedness, not style
Gate
Invented citation is caught
The instrument
If two humans disagree, you are not measuring the model. You are measuring a foggy guide. Fix the page first.
Pay for hard
Speed invents yes. Pay for careful hard cases. Three labels only: grounded, unsupported, refuse.
Worked example
Two red pens
Two raters, one lie, two different marks. The model is innocent. The guide is broken.
Picture to keep
Guide · Example · Agreement
- 1Three labels only
- 2One page
- 3Two humans, one lie
