Research Partnerships
Custom evaluations, data collection, and tooling for AI research teams.
We collaborate on scoped studies and longer-term data programs. Initial work can be a small evaluation or collection pilot, with the protocol, quality bar, timeline, and permitted use agreed together.
Model evaluation studies
Studies of output preference and task performance across models, checkpoints, and agent configurations. The protocol follows the research question, including the task distribution, comparison conditions, and review criteria.
- Candidate and reference model comparisons
- Task selection, rubrics, and human review
- Preference analysis and representative examples
Custom data collection
Collections designed around a capability, domain, or observed failure mode. We work with the research team on the specification, collection workflow, and acceptance criteria.
- Prompts, source inputs, and task contexts
- Preference pairs, explanations, and trajectories
- Domain requirements, review, and delivery formats
Research tooling & environments
Purpose-built infrastructure for collecting and evaluating agent work. An engagement can include task environments, review interfaces, or an ongoing collection pipeline.
- API, MCP, browser, and software environments
- Task verifiers and annotation workflows
- Iterative collection and dataset delivery
Research inquiries
Evaluation studies, custom collections, and research partnerships.