Traverse Captures Human Judgment to Train AI on the Subjective Stuff

The Macro: AI Has a Taste Problem

AI models are very good at tasks with objectively correct answers. Code that compiles or does not. Math that is right or wrong. Facts that are true or false. But most white-collar work is not like that. A lawyer drafting a contract makes judgment calls about risk tolerance. A doctor interpreting symptoms weighs competing hypotheses. A writer choosing between two phrasings relies on taste. A salesperson deciding when to push and when to back off reads social cues.

This subjective, taste-dependent work is the hardest unsolved problem in AI training. You cannot generate training data for it synthetically because there is no ground truth to optimize against. You cannot label it cheaply because it requires domain expertise. And the existing approaches, RLHF with generic human preferences, produce models that are average at everything instead of excellent at anything specific.

The frontier labs know this is the bottleneck. The models are big enough. The compute is available. What is missing is high-quality training data for domains where the right answer depends on context, experience, and judgment. Someone needs to build the infrastructure to capture and structure this kind of data at scale.

The Micro: Controlled Environments for Capturing Judgment

Lance Yan and Zachary Yu founded Traverse. Lance is CEO, focused on “turning profit.” Zachary is CTO, focused on “solving subjectivity.” They are a two-person team from San Francisco, part of YC Winter 2026 with Jared Friedman. Their angel investors come from OpenAI, DeepMind, Anthropic, and Meta, which tells you that the people closest to the training data problem think Traverse is working on the right thing.

Traverse captures human behavior in controlled environments to create training data for AI models in subjective domains. Instead of asking people to label data after the fact, Traverse designs environments where domain experts make real decisions, and the decision-making process gets captured in structured form.

Think of it like a flight simulator for knowledge work. A lawyer works through a contract negotiation in a controlled setting. Every decision, every trade-off, every judgment call gets recorded with the reasoning behind it. That structured decision data becomes training material for models that need to develop judgment in legal work.

The company partners with frontier AI labs to produce this data. The focus on law, healthcare, sales, and writing suggests these are the initial domains where they have built expertise.

The Verdict

Traverse is working on one of the most important problems in AI. Training data quality is the binding constraint on model capability, and the hardest training data to produce is for subjective, judgment-heavy work. The angel investors from all four major AI labs validate the problem statement.

The risk is that the frontier labs build this capability internally. If a major lab decides that capturing human judgment is a core competency, they have the resources to do it at massive scale. Traverse needs to build domain expertise and data capture methodologies that are hard to replicate.

The competitive set is sparse. Scale AI does general data labeling. Invisible AI does process documentation. But nobody is specifically focused on capturing subjective human judgment as training data. Traverse has a real opportunity to define this category.

In 30 days, I want to see the list of lab partnerships. In 60 days, the question is whether the training data Traverse produces measurably improves model performance on subjective tasks. In 90 days, I want to know about domain expansion. How many different professional domains is Traverse capturing judgment data for? The broader the coverage, the more valuable the platform.