Every result record the harness writes is otherwise an island. plan_d/src/graph.py projects all of them into one graph — 201 nodes and 632 edges over models, tasks, datasets, modalities, held-out generations and acquisition settings — and this page walks that projection backwards from a single number. Each card names the file or the edge its facts were read from, so nothing above has to be taken on trust: every line is checkable against a path in the repo.
The gaps are load-bearing. 5 node types are declared and empty — hospital, deployment, scanner, clinician, publication — and where a chain needs one, it renders the hole and the graph’s own reason for it rather than stepping over it quietly. A provenance walk that hides what it cannot answer is worth less than one that stops.
Nothing here re-computes a score and nothing here checks a signature. The chain ends at a signed document and hands you to Check a certificate, which does.
Trust Graph