Research

What we are finding out

Experiments on execution verified environments, with the methodology and the limitations attached. Every number here is reproducible from the public dataset.

·7 min read

Training a 14B to patch authorization bugs

Fine-tuning Qwen3-14B on VibeSec takes it from 6.0% to 78.0% secure patches on held-out tasks, a +72 point move. How the run was set up, how it is graded, and what the number does not show.

Read it