Watch the idea
The concept
Case
Question + allowed sources
Score
Groundedness, not style
Gate
Invented citation is caught
LLM-as-judge is a tool, not a court
A judge model can be sweet to fluent lies. Humans still own the gold.
Then prove it
Open the lab. Pass it. Write one sentence you could say in an interview.
Worked example
The artefact
A judge model can be sweet to fluent lies. Humans still own the gold.
Picture to keep
Cheap pass · Human gold · No self-grade
- 1Judge is a sieve
- 2Humans are the court
- 3Different model if you must
