Healthcare AI Citation UX Review Kit: Six Evidence Checks Before Development

Abstract

Healthcare AI Citation UX Review Kit gives product leaders, CTOs and design leads six evidence checks to use before commissioning or extending a healthcare AI knowledge product. This software deposit preserves the interactive worksheet, its review catalog and reproducible Markdown and JSON exports. A team can use it during a prototype review to identify the evidence it still needs from a delivery partner. The starting point is a concrete design failure. In its own aesthetic-medicine learning-platform mockups, Pharos Production reported that 6 of 7 demo citations were wrong, including 4 invented citations. The project was still in its design phase. The healthcare AI citation UX case study explains what the team changed and which checks remained proposed work. Those observations concern one internal design audit, not a measured failure rate for healthcare AI systems. What to bring to a prototype review Ask the team to demonstrate one claim from the answer screen through to its supporting passage. Then record what you actually inspected. The six checks cover claim support, source location and edition, visible separation of sample content, missing-evidence states, distinct source categories and changes that must survive handoff. Each check names an artifact to request, an action to observe and a failure to watch for. An accessible link alone cannot establish that a passage supports the displayed claim. Likewise, a missing source and a failed retrieval are different conditions. The worksheet turns these distinctions into review questions that a product owner can discuss with engineering and design before accepting the next scope of work. A brief your team can use The interface accepts three self-reported statuses: Not reviewed, Gap found and Documented. The exported brief places gaps before unreviewed items and retains all six checks, including documented items. Selecting Documented records a person's assessment. Even when every item has that status, the output still asks the team to review the evidence. Use the interactive citation review worksheet or run the archived static application. The deposit also contains an editable buyer review sheet and two labeled example exports: an untouched review and an illustrative mixed review. The examples contain no client findings. No registration is needed to download the files or try the public worksheet. Browser selections are not persisted or submitted to a server by the application. Version and reproduction The archive freezes kit version 2026-09-26.1 at source commit be3937d2f4e8307160a636439d5888bf05f9ccb5. The source package's internal npm version is 1.0.0; the dated kit version identifies the review catalog and exported briefs. Run npm run verify from the source directory to execute the six contract tests and check the archived static-file hashes. The code uses Node.js built-ins and has no package dependencies. Serve the source/docs directory over HTTP to run the interface. This release preserves the source files without changing their original GitHub Pages campaign links. The links in this Zenodo description and its accompanying buyer guide use a separate Zenodo referral campaign. Checksums and a source manifest distinguish the preserved implementation from the added release documentation. From evidence gaps to a development conversation Pharos Production, the software development company behind the kit, connects these review questions to its documented healthcare product discovery and UX work. After recording your evidence gaps, read the healthcare AI prototype design and citation audit. Compare its source-verification and handoff decisions with the work you need a delivery team to demonstrate. A useful discussion starts with the unresolved evidence requests in your brief. The six checks are an editorial selection. The kit does not inspect medical evidence, test a model, approve a vendor or establish clinical accuracy, regulatory conformity or readiness for deployment. Research and implementation were AI-assisted. The code is released under MIT; review content and release documentation use CC BY 4.0. Linked third-party sources retain their own rights.