How MndUP (mndup.com) is measured against your own decisions
MndUP (mndup.com) is measured against your organisation's own earlier decisions, not against a benchmark of our choosing. This page describes the method; it publishes no result.
What is compared
Your organisation keeps a reference set: documents with the outcome your assessors reached. A check runs the same screening on those documents and compares its results with yours: how often they agree, agreement measures that allow for chance, the rank order, and how many of the applications you funded it places above the line.
The hold-out
Each document in the set is marked for tuning or held out. A document whose answer was shown while the screening was being tuned counts as tuning. Results are reported on the held-out documents only, so the screening is not graded on answers it has already seen.
The spread between two unchanged runs
AI models do not answer exactly the same way twice. So the check also runs the unchanged screening twice and measures how far its results move by chance alone. A change only counts as better or worse when the difference is larger than that spread and a paired interval around the difference stays clear of zero.
Where the expected answers stay
The expected outcomes stay with your organisation's own screening. They never go into a prompt, an export or another organisation.
What we do not claim
We publish no accuracy figure. Results depend on your rules, your documents and the models you choose, so the only number that means anything is the one measured on your own decisions. Any example figure on this site is synthetic and says so.