rmerk closes the peer-review loophole in med-mal extraction
A compliance gate that only read typed text now reads the scans too - and there is no off switch.
The original prescan was supposed to halt extraction whenever a medical record carried peer-review or M&M conference markers, which in most US jurisdictions are legally protected from discovery. The catch: it only checked the typed text layer, so scanned pages where those phrases appeared as pixels could slip through and end up as extracted event rows tied to protected material.
rmerk's fix renders the image of any page without searchable text and asks Claude whether the protected markers are visible before any event extraction runs. There is deliberately no environment flag to disable it, even though on a 3,000-page hospital chart that's $5-7 in image-recognition calls before the real work begins. The author frames the cost as the price of a gate that can't be quietly turned off in production. Sitting underneath is a broader rebuild of the extraction pipeline, with new database migrations, an optional async queue mode for serverless deployments, and proper test coverage.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?