Disclosure of performance evidence in FDA decision documents for AI-enabled devices before and after the 2025 draft guidance: data and code

Reproducibility package for a protocol-frozen interrupted time-series study of 977 US FDA decision documents for AI-enabled medical devices, comparing public disclosure of performance evidence before and after the FDA’s 7 January 2025 draft guidance on AI-enabled device software functions. Contents: the FDA AI-Enabled Medical Device List as retrieved (3 November 2026); a document manifest with source URLs and SHA-256 hashes; page-marked text of all coded documents; all coding passes with verbatim quotations; adjudication records; frozen datasets; protocol, codebook, deviation log and all agent instructions; analysis code, outputs and figures; and a human-validation kit. Running run_all.sh reproduces every reported estimate. Important: document coding was performed by large language model agents (Claude, Anthropic) without human verification. Agreement between coding passes measures repeatability, not accuracy. See README.md for details and limitations.