sean

Scores a Claude Code harness from the transcripts already on your machine: 6 stats of what your agent actually did, computed locally with no LLM calls. Then it runs a reproducibility gate over its own numbers and prints which ones failed and which it could not test at all.