Skip to main content
Eval Labs must make reviewing Lucia feel simple, calm, and low-pressure because reviewer fatigue damages evaluation quality.

​
Doctrine

Reviewer cognitive load is not a UI detail. It is a data-quality risk. If employees feel confused, intimidated, or forced to think like AI experts, the review signal becomes weaker.

​
What reviewers should not have to do

Reviewers should not have to:
  • understand AI training theory
  • invent intent labels
  • decide taxonomy names
  • write long analytical notes
  • explain model behavior
  • know what adjudication metadata means

​
What reviewers should do

Reviewers should:
  • read the prompt
  • read Lucia’s response
  • use guided controls
  • answer quickly and honestly
  • flag senior review when uncertain
  • add a short note only when it helps

​
Interface rule

Eval Labs should prefer: 
guided choices
over
open-ended interpretation
and: 
visual cognition
over
form-filling

​
Why semantic UI matters

Semantic controls such as confidence sliders reduce translation burden. The reviewer should feel the difference between: 
weak / concerning
uncertain / mixed
strong / confident
without needing to reread a scoring guide every time.

​
Canon rule

A review interface that creates psychological paralysis will produce worse training signal, even if the schema is technically correct.