A map you can trust — because it shows its work

This site began as the working group's own document — the questions and findings the group assembled to prepare for its meetings with Anthropic. At a working-group meeting, that document was used, live, to charter a genesis campaign for a context graph: a structured knowledge base integrating the context behind every topic and question in the working document — references, examples, and case studies — with a chip on every claim saying how it was checked and how far to trust it. An agent working in the aDNA format built the first version of that graph, and the site you are reading from it, in about thirty minutes, during the meeting itself.

What it teaches is the material the group gathered about using Anthropic's Claude safely in rare-disease advocacy: folder access, secrets, data retention, HIPAA, Skills, and the questions still open. It is built for people doing patient-facing work who did not build these tools — plain language first, detail underneath, provenance on everything.

Two ways in. Advocacy and operations staff: read straight down — the six modules below are in teaching order. Clinicians and clinical researchers: start with the clinician path, which takes the same findings in a different order — what is covered, what is retained, and what the tool must not be trusted with. Either way you are reading the same graph, with the same chips and the same receipts. There is also a printable summary built to be handed to someone — the rules, with their dates and their marks, on paper.

Why this exists (and why it looks the way it does)

Your document asked: “How can we control what sources Claude pulls from?” One working answer is a context graph — a curated set of fact files, one claim per file, each carrying its source, its retrieval date, and an honest confidence rating, that an AI assistant is instructed to answer from, instead of from its general memory. This site is that answer, demonstrated: every page you're reading is rendered from exactly such a graph, and the graph itself is what the group's assistant queries. Control the graph, and you control the sources.

So this site is two things at once. It is the body of knowledge — the findings, rules, open questions, and validation state that medical and advocacy leaders need to use these tools for the good of patients. And it is a demonstration: the first version of the graph behind it was built from the group's working document in half an hour, by an agent, with every claim carrying its provenance — and what you are reading now has been through the review rounds recorded in the verification log since. The chips are not decoration — knowing how each piece of information was validated, and to what degree, is a critical dimension of building and understanding a system like this. What took thirty minutes was that first version — and that is what these tools now make possible for any group willing to show its work.

How to read the chips

Every claim carries a chip that tells you how much to lean on it. Each state is marked three ways at once — colour, icon, and label — so it reads the same in colour, in greyscale, and to a screen reader. This legend is on every page of the site; the printable one-pager carries the same marks in its own compact key, sized for paper.

The six modules

The call agenda

Every open question across all six modules, consolidated and printable: the Open-Questions Register — the working group's agenda for its Anthropic contact. Currency baseline: 32 of 44 findings confirmed against primary sources, last verified 2026-09-19.