Same content, two shapes. The left is how the knowledge is stored. The right is how it gets used. Extracting the second from the first is the compression — and it is also what makes the reasoning teachable, because a trainee cannot reason from prose they were never given the skeleton of.
The gap report is not an apology — it is the working list. These are the points where the sources disagree, or where the text assumes something it never states. Each one is a question for the clinician who owns the framework, and answering them is what turns an extraction into a teaching object.
Why not simply chunk the books and retrieve by
similarity. That build demonstrates well and disappoints on real
questions. Clinical reasoning is long-form and conditional: a framework
spans pages, and chunking severs the sequence, returning a fragment that
reads authoritative while having lost the preconditions that made it
correct. Plausible advice detached from when it applies is worse than no
answer. Structure is extracted first; retrieval then runs over the
structure, with the source text cited underneath.
What is shell and what is content. The viewer, the graph
model, the citation requirement and the gap report are engine and carry no
clinical claim. A framework is an assertion about how to reason, and must be
extracted from a named source and approved by a named clinician. The example
above is built from published consensus definitions to show the shape;
nobody has verified it, and the interface says so rather than hiding it.