Screening as a Modeling Choice: How Deterministic and System-One Semantic Gates Shape Reconstructed Co-authorship Networks

Co-authorship networks built from institutional web pages are exposed to the risk of topological distortion and error propagation due to the heterogeneous data contained within HTML code and the presence of unstructured strings. In settings where large volumes of documentary records make exhaustive manual verification impractical, automated screening becomes necessary, and the rules used to select records may influence the resulting network topology. In this paper, we present two different approaches to address this problem. The first is the traditional approach that uses complex regular expressions to identify co-authorship strings (Method A); the second, instead, uses a new family of models called “System-One” models. This model allows us to analyze a text string and categorize it with the capabilities of generative AI, but without generating a single token (Method B). The model is useful for its capability to output probabilities across a fixed set of choices—in our case, TRUE, FALSE, and INSUFFICIENT—based on a set of instructions. To control how the model categorizes each string, we subsequently introduce a new policy called Method B-selective to isolate low-confidence or ambiguous decisions: |pT − pF | < 0.20 ∨ pI > 0.15. This analysis was conducted on 1,848 total bibliographic strings extracted from three university department websites. OpenAlex API services were implemented to resolve individual identities through the OpenAlex-ID, ORCID, and the profile webpage of each individual. The impact of the screening process depends on the HTML structure of the department: in PPA, where lists intermix publications, awards, and editorial work, Method A includes 87 single-authored papers (out of 166), whereas B-native filters out the noise—reducing the number of resolved papers to 98 while increasing the number of edges from 237 to 247. Furthermore, fixed-identity diagnostics reveal local fractures (one component split and 15 disconnected author pairs) that remain invisible to global graph metrics. Upstream screening is a substantive modeling choice, and typed semantic interfaces with explicit abstention offer a reproducible paradigm for Social Network Analysis based on heterogeneous web data.