Methods

Corpus

The essay draws on over 2,000 public texts about open science. The collection inventory lists 2,098 collected texts by title, type, document ID, and text hash. The inventory combines 2,031 texts exported from a frozen collection with 67 texts from sources gathered for close reading. Claim extraction covered the 153 documents selected for analysis from that inventory. They include declarations, recommendations, policies, consultation responses, research papers, and infrastructure principles. One additional selected document was excluded because its OCR text was unreliable. The manifest identifies each analyzed document and records its source, text hash, and claim count.

Claim extraction

Claude Sonnet 5 extracted 21,422 claims using the extraction prompt and schema. Long documents were split at section boundaries. Each claim retains its source passage and fields describing its kind, modality, actor, and subject. The prompt admits general empirical, normative, and definitional statements; it excludes commitments, reporting procedures, and particular events or statistics. Passage-location checks tested whether the excerpts occurred in the source text.

Comparison

Embedding similarity identified cross-document candidate pairs with cosine similarity of at least 0.88. The equivalence rubric tested whether each pair expressed the same proposition. Grouping required agreement across repeated judgments and excluded incompatible links. This produced 2,517 claim groups containing 6,917 claims from 150 documents; 14,505 claims remained ungrouped. Each group spans at least two documents.

Two approaches organized these groups. The theme classifier first assigned groups to essay-derived themes, then organized the remaining groups into additional themes. Across successive layers, it defined 71 themes and assigned groups to 69 of them, accepting assignments when both judgments agreed and neither had low confidence. Separately, embedding-based clustering produced 67 clusters, described with a naming prompt. The resulting counts include each document once within a theme or cluster.

The contradiction test examined 3,859 candidate pairs of claim groups. Five pairs were retained as direct contradictions after repeated judgments and review, all concerning hybrid-journal payments. The results identify the opposed groups; these five contradictions are supplied separately from the claim graph.

Counting rules and limitations

The counts in the essay use all 153 analyzed documents as their denominator. A document counts once within a theme if at least one of its claims belongs to a retained group assigned to that theme. Each group contains claims judged equivalent across at least two documents. The groups contain claims from 150 documents; three other analyzed documents have extracted claims but no retained group membership.

Theme membership can reflect a complete proposition, a component, a premise, or either side of a disagreement. The counts describe saved assignments, not validated endorsements of the complete subsection headings. Theme definitions differ in breadth, so their totals do not rank the strength of agreement.

The counts retain every analyzed document as a separate record. Related versions and organizational responses do not represent independent constituencies.

In sections 1 and 2, the bullets under each heading list the theme’s claim groups that the most analyzed documents contain, with each group’s document count. A group is eligible only if its documents come from at least three different works (two in section 2, where positions divide) and no single work supplies more than half of them. Versions, translations, and series of one work count as one work, so repeated filings or revisions cannot fill the list. Claims tied to one country or institution, and near-duplicates of a listed claim, were skipped. Each bullet paraphrases its group’s representative claim. The major-documents line names well-known documents that contribute claim groups to the theme. essay-claim-bullets.json records the groups, themes, and works behind these numbers, and the publication build checks each number against the data.

These model-assisted classifications describe a selected corpus. Candidate selection and classification errors can miss or misgroup claims, so absent matches do not establish absence of a position. The two grouping approaches share their underlying claims and judgments.