Traverse

Traverse is a research lab that partners with frontier AI labs to produce training data enabling models to develop taste, judgment, and decision-making for non-verifiable tasks. Unlike traditional approaches targeting deterministic domains like math and coding, Traverse focuses on inherently ambiguous work found in law, healthcare, sales, writing, and strategic decision-making. They believe non-deterministic work becomes verifiable when enough contextual information is captured about the situation, constraints, and reasoning behind expert decisions. By observing real experts in real environments and preserving their reasoning processes, Traverse creates training signals that allow models to learn how judgment actually works. Their approach involves capturing reasoning processes behind expert decisions at scale, producing a class of data that currently doesn't exist. They aim to give frontier models the foundations to eventually surpass human white-collar work, accelerating the path toward superintelligence. Key features include RL environments for non-verifiable tasks, training data production for models to develop taste and judgment, focus on subjective long-horizon tasks requiring reasoning and context, and partnerships with frontier AI labs to generate training signals. Traverse operates as a research lab, not a product company, meaning there is no public API or ready-to-use tool. Compared to competitors focused on deterministic RL (like math and coding environments), Traverse is pioneering training methods for ambiguous, economically valuable domains. It's solely relevant for frontier AI labs capable of partnering on long-term research; individual developers or small teams without those resources will not find a usable product here.