Harness Experiments

A harness is everything around the model: its tools, system prompt, edit format, error handling and context management. Most comparisons swap whole products at once, so you learn which bundle won but not why.

These experiments change one part of a harness at a time, run the same tasks across several model families, and grade every run with a script rather than an opinion. Each result reports success rate, cost per solved task and how runs failed, with the uncertainty stated.

The first study is being designed now. Results, methods and raw data will be published here.