For AI that is already doing real work
Review what your AI does in production.
Once it is live, people review a sample Oloproof draws from real traffic, in the browser, while the content stays on your side. The workspace decides on what they found, not on what anyone chose to show it.
Draw
A sample nobody chose
Your collector receives your traces over OpenTelemetry and commits to every hour it sealed with a signed Merkle root. The workspace then draws the sample with its own randomness and checks each drawn trace against that commitment, so no trace the collector sealed can be held back from the draw.
Review
Reviewers in the browser
An Owner or Admin opens a review queue and assigns reviewers. Each gives a pass or fail verdict, a score on a declared scale, or a blind preference between two outputs, recorded under their own account. Items drawn for measurement are shown without the judge's verdict.
Keep
Content stays on your side
Under the default egress policy the workspace holds no inputs or outputs. The reviewer's browser fetches each item's content from the collector on your network, on a five-minute ticket the deployment signs, and Oloproof never stores it.
What the workspace decides
Pass or fail verdicts on a sample the workspace drew correct a judge's pass rate, by prediction-powered inference, and the workspace verifies that interval with its own engine. It counts only labels it sealed itself, from independent accounts, and decides with the same four states as every run: pass, fail, insufficient evidence or manual review. Scores and preferences are stored and shown, and enter no statistic until a method for them is admitted.