Bounded Verification and Asymmetric Recoverability in Biomedical Knowledge-Graph Construction

Automated construction of scientific knowledge graphs can fail even when individual model judgments appear reasonable: an early filtering error can remove evidence from all subsequent observation. This creates an asymmetry between false retention, which preserves a candidate for later correction, and false deletion, which can make that candidate unavailable to downstream audit. We present an evidence-compilation architecture organized around this asymmetry. Entity identities and admissible relation menus are fixed before semantic adjudication; models make bounded contextual decisions, compiler checks constrain what can enter the graph, and consequential rejections are retained as audit data. Destructive filtering policies are evaluated on their own false-drop populations, and policies whose measured risk is unacceptable are disabled or diverted instead of being treated as trustworthy by default. For a post-hoc relation verifier, we instantiate this principle with a relation-specific treatment based on the Wilson upper bound. A pre-registered benchmark further tests whether claimed support exists in the cited source, including a zero-co-occurrence stratum on which the graph can refuse when supporting evidence is absent. Each of the 162,155 served relationships retains its supporting evidence and versioned construction metadata. The resulting design treats verification as a property of the compilation process, not solely of the model producing an individual judgment.