Measured output quality for an LLM pipeline

A publication gate for generated content. The pipeline produces text, an automated judge scores it, and anything below the bar never reaches the client.

Quality is a number here, not an opinion: a 379-example reference dataset, macro-F1 0.81, LLM-as-a-judge scoring. The gate stopped 7 runs out of 31 before anything was published. Median run cost $0.13 and took 12.8 minutes.

Stack: Python, LangChain, OpenRouter model routing, PostgreSQL.