01
Extraction
Was the source read correctly?
ONGOING
Measuring where automation succeeds and fails.
We evaluate extraction accuracy, hallucination rate, provenance accuracy, labor savings, and the share of assertions that require correction.
/work/ai-research-evaluationCultureOS Labs · 2026
MEASURES
Accuracy is not one number.
02
Reasoning
Does the conclusion follow from the evidence?
03
Review
What did a scholar correct, reject, or investigate?
DISCIPLINE
Failure cases are results.
We publish where automation failed, where uncertainty remained, and where human review changed the outcome.
START HERE
What is your institution carrying?
We start with the one workflow costing you the most.