Research · public studies

Research threads

Deep dives and learnings from the lab on trajectory tooling, harnesses, and reporting discipline for RLVR stacks.

← Back to lab overview

Cheap Reward Hacking Detection

A 14M-parameter encoder matches a frontier LLM judge at detecting reward hacking in agent trajectories — at four orders of magnitude lower cost.

Open publication →

Coding Harness Inspection

Analyzing the adoption, market share, and configuration patterns of AI coding harnesses by examining approximately 400,000 public GitHub repositories to serve as a diagnostic thermometer for the evolving ecosystem.

Open publication →