Abstract
<title>Abstract</title> <p>A knowledge base earns its value less from the facts it stores than from the connections it can compose across them — connections that no single source states. Composing such links, however, risks fabrication, so the property that matters is not that links are surfaced but that they are correct. We report an independently validated test of cross-source synthesis by a proprietary engine operating over a compiled medical knowledge base. The behavioural claim, fixed in advance, is that given an entity the engine surfaces structural connections that are not stated in any single ingested source and are independently verifiable against external literature the engine did not use. Because the production setting — patient data — cannot be exposed, the capability was validated on a checkable surrogate, textbook medicine, with the operation asserted identical and only the input differing. Validation was performed by an architecturally independent check against published literature; the engine did not grade its own output. A representative sample of nine composed links, spanning mechanistic, clinical, and knowledge-to-compute domains and each drawn from multiple independent sources, was examined: nine of nine were confirmed correct. A discrimination control — a real cross-organ pattern the engine recognised but deliberately withheld — showed the synthesis is selective, not indiscriminate. The links are correct, not novel: established literature is the validation instrument, which is what shows the composition is right rather than hallucinated. The capability was demonstrated at an early stage of the knowledge base's compilation, well below its intended scale. The claim is reliable cross-source synthesis, not discovery; validated synthesis of this kind is the precondition for computational discovery, which we frame as forward outlook, not a result.</p>