Methods and data

Methods & data

This guide explains which records were included, how the figures were calculated, and what was checked.

Complete record accounting: Article coverage and credit evidence separates source-supported, declared-only and unresolved account credits. Reviewed does not mean personally confirmed. The published coverage ledger shows inclusion and credit evidence.

1. What is included?

The archive dataset contains the 423 published Tagging the Tower records. The source export also contains nine category drafts, excluded from the dataset. Of the published records, 327 have no detected external-curation marker and 96 are curated/source-linked. Local/unflagged is a detection category, not proof of authorship. The NLTK/MALLET corpus includes 301 published posts after further documented exclusions.

Current audited source inventory · Additional grouped-name view · Model inclusion/exclusion ledger · Data definitions and CSV schema index

Plain-language glossary

Corpus manifest: The per-post ledger saying which texts were included and why others were excluded.
Collocation: A word pair that occurs together in the cleaned texts; its score is not proof of a shared concept.
Concordance: Examples of a word in its surrounding text, used to inspect context.
Representative post: A post selected for inspection, often because of a high topic weight; selection does not guarantee that it represents the whole archive.
Theme coding: Assigning interpretive categories to texts. In this project, a coding file is not evidence that an independent human panel assigned or confirmed its categories.
Academic year: August–July here, unlike the January–December calendar year. Fall 2026 belongs to 2026–2027.
Vocabulary exclusions: Terms removed by explicit cleaning or frequency rules. Their absence from a model does not mean their ideas were absent from the source writing.

2. How were words and topics produced?

A corpus is the set of texts analyzed. Stopwords are common or project-specific terms removed by the saved cleaning rules. A topic model (LDA) estimates recurring word groupings; its weights are statistical descriptions, not verified meanings. k is the chosen number of topics. MALLET fitted these models; NLTK supplied language processing, word counts and phrase statistics.

The primary display uses 8 topics, seed 42. It is not declared optimal: compare k=6/8/10 and other seeds. A seed changes the model’s initial starting point; seed stability checks whether a pattern persists. Author sensitivity asks how aggregates change when a credited writer’s posts are removed; it is not a causal estimate of that person’s influence.

Full setup and model settings · Custom stopwords · Vocabulary exclusions · Model comparison · Seed stability · Author sensitivity

3. Who wrote the posts, and who was a Fellow?

A byline names a writer in a post. A byline candidate is a search result requiring review, not an accepted credit. A merged author entry groups documented name variants while preserving original strings. A cohort often means an entering group; this project instead has incomplete signed lists of active Fellows for particular semesters. Publication credits do not establish appointments, and a missing match does not establish nonmembership.

November 2 source-credit correction: Ten additional records have corrected names, including McMichael’s linked-data posts. Reviewed author fields and institutional-guide evidence now qualify for the reviewed-credit filter. Account-only credits remain provisional. Ten corrections · November 2 unresolved-credit snapshot.

November 2 attribution audit · Identity decisions · Signed-list evidence · Source writers and CRediT/GAIDeT contributions

The archive also includes guest-authored posts by students who were not Fellows at publication, including participants in the Graduate Center Digital Research Institute (GCDRI). Ian supplied this contextual clarification. Guest authorship is an authorship relationship; participation and fellowship appointments are separate roles. Assess fellowship status at the post date, rather than treating it as a lifetime identity. A missing directory match alone establishes neither guest status nor nonmembership. Participant reflections remain in the full archive and its applicable analytical populations; a Fellow-only filter intentionally selects a narrower population.

Historical sources

Annual report comparison and source tables document reported work and staffing. Archived website source records provide independent context. The reviewed credits link the original archived article captures used to restore 43 bylines. Report staffing and project credit are separate from article authorship.

Themes and connections

The thematic visualizations compare eight saved keyword-signal categories across 327 published local/unflagged posts. A separate layer shows the essay’s six interpretive themes only for its 19 selected examples. These layers are not MALLET topics or independent human coding. Yearly shares include unmatched posts in their denominator; co-occurrence counts the intersection of post sets. Essay-sample counts do not establish archive-wide prevalence.

Thematic method · Unfiltered annual thematic counts · Assignments and dictionary definitions

4. What kind of review has happened?

Automated checks compare saved calculations and protect the baseline. Agent-run source checks verify literal excerpts. Browser checks exercise navigation. The two Codex sub-agent reviews and the supplied external persona review are simulations, not actual user studies. Ian identifies the external model as Deepseek V4.1 Flash 1017, running in the CAIL sandbox in OpenWebUI on a ZIP of an earlier iteration. The external document reports manifest-level inspection; the exact input ZIP is not recorded. Codex checked its recommendations against the current sources before implementing changes. Files named “theme coding” or “interpretation review” do not by themselves establish independent human coding. The current topic labels and prose are agent-developed; complete human confirmation is not documented.

Work log, contributor roles, and release checks · Reviewer guide · Interpretive source trail

5. How can this be reproduced or reused?

Opening index.html needs no installation. The package README gives three commands to serve and check the saved package, and separately explains the external dependencies for a full NLP refit. Such a refit must be kept as a new analysis version: historical byline handling and credits have evolved. Full retraining runtime is not benchmarked here.

Rights status and suggested citation · Accessibility · Submit a correction