Internalization as an Operation: Recovering the Dependency Structure of a Result, and a Pre-Registered Failure to Do So

Abstract

Machine systems now produce results faster than readers absorb them, and machine-generated summaries are widely observed to emphasize the wrong material. This paper argues the failure is structural, not one of model capability. Given only prose, a renderer can weight content by prose mass — the text an idea occupies — while the property it needs is structural load, defined over the dependency graph. The two diverge: a one-line lemma can carry an argument that pages of routine verification do not. The paper specifies a five-step internalization procedure, the inverse of structured drafting — prose to dependency graph to a second rendering — and supplies the measurement it requires: a miracle count, the number of steps verified but not derived from what a declared reader holds. Decision rules and agreement thresholds were fixed before any run. In the pre-registered validation, two machine operators from different model families extracted five documents; edge-level agreement fell below the declared threshold on every one, and the operators diverged on which nodes exist before how nodes connect. That extraction is well posed is refuted at that level, no downstream quantity is licensed, and the specification, the instrument, and the negative result are reported together. Includes zharnikov-2026bk-internalization-as-operation.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.