Rox published the Rox Outbound Reward Model, a model built with hiddenweights that scores cold emails, trained on drafts before and after SDR edits and tuned on open and reply data. On 717 held out SDR edits it preferred the edited version 80% of the time, against below 50% for GPT and Claude used as judges, and the emails it scored in the top 20% drew 1.75 times the average reply rate. Rox says it uses the model to pick the best of many drafts in its Autonomous Outbound Worker.
Rox is disclosing how its agent chooses which email to send and publishing test results for it, which is more model transparency than most AI SDR vendors offer.