2201344667392187

People who use screen readers can miss what a chart shows or hear a document's heading after its instructions. Public bodies, universities and healthcare publishers must make their documents accessible: covered US state and local government content by 26 April 2027 or 2028 depending on size, and covered HHS funding recipients with at least 15 employees by 11 May 2027. Docket checks PDFs, repairs what it can repair safely and gives a reviewer checked drafts for the decisions that need judgment. During the hackathon, Meta's Muse Code built the module that lets Muse Spark draft figure descriptions grounded in the document and choose a reading order only from the options Docket's engine offers. We sent 391 figures from a 109-document test corpus to Muse Spark: every call completed, for about $1.89 at our configured prices, typically in under three seconds, and every description it wrote repeated only text the picture itself shows. On eleven reading-order choices it agreed with the founder ten times; its two proposed fixes save eleven manual moves, and a person approved the heading-first order in a live rehearsal. With the Meta routes switched off, the build gives the same results as before the hackathon: 154 of 154 verdicts and 7 of 7 output files byte for byte. Nothing is applied without a person. Built for the Meta Global AI Developer Hackathon, startup track, Build for Your World. Built before the hackathon: Docket's platform, engine, workbench, evidence pack, forms, inventory and family memory. Built during the hackathon: the Meta provider route, the tenant route selection, the reading-order choice among engine candidates, the measurement drivers, the SAM experiment code (not run live), their tests and measurements, the public repository, the write-up and the video. In the repository, only java/muse-code-hackathon/src contains Docket application source; the existing platform is supplied as compiled binaries so the module builds and the product runs.