How fast can a coding agent p-hack? Measured: on known-zero data, realistic searches manufacture p<.05 on 33-97% of null draws, in ~1 second each. Instrumented specification-search engine + 11 Claude Code skills for econometric designs; every search leaves a verifiable ledger and an honest p-value. Research and teaching use only.
Benchmark for statistically valid AI scientist systems, using audit-closed protocols, transparency logs, and sequential inference to prevent false discoveries in autonomous research agents.
