Unitization Before Extraction: A Pre-Registered Decomposition of Disagreement in Machine Recovery of Dependency Structure

Abstract

Machine-extracted argument structures are increasingly read as evidence about how fields move, so measurement error entering at extraction propagates into whatever is built on the extracted graph. In a prior pre-registered run, two machine operators did not recover the dependency structure of five argumentative documents at a declared agreement level, and the failure sat upstream of relation labelling. That result cannot say what the disagreement is made of: unitization, node typing and relation drawing varied at once, and each document was extracted once per operator, so between-operator disagreement is not separable from run-to-run variance. This paper pre-registers a design removing both. A unit inventory is fixed before extraction and supplied identically to both operators, making node selection a decision over a common index set; every extraction is repeated, supplying a within-operator baseline against which any between-operator number is read. Three unitization conditions run on one corpus, one operator pair, one epoch, ordered in advance. Thresholds are re-derived per layer from their own reference classes, not inherited, and the sensitivity table is published empty. The claim under test is the predecessor's closing recommendation, stated so it can fail: if fixing the units does not move edge-level agreement, the failure was mislocated. Includes zharnikov-2026bl-unitization-before-extraction.yaml (Paper Spec v0.1.0) — a machine-readable specification of the paper’s claims, assumptions, and dependencies. The paper’s full machine-first bundle (the SPINE claim/dependency graph and the ONTOLOGY term module) lives in the public repository; see github.com/spectralbranding/paper-spec for the standard. This PDF is generated programmatically from that machine-first source under a research-as-repository model.