Search posts
Look up a writer, tool, or topic in the full text.
Loading the complete saved reading text. Search will be ready when all records are available.
Research prototype
Learn from over 14 years of doctoral student contributions.
Look up a writer, tool, or topic in the full text.
Open one of the 423 published records at random.
See what 21 selected posts suggest about digital scholarship.
Search whole words in any order, or an exact phrase, across full saved text. Case, accents and punctuation are ignored.
Each collection draws an eligible article at random, then one sentence. The broader collection requires source-supported credits; neither covers all 423 records.
Loading an excerpt…
The starting collection contains 30 exact excerpts from 25 articles, selected with AI assistance at Ian G. Williams’s direction for curiosity, recognition, emotional connection and meaningful ideas. Joy, humor, uncertainty, care and unfinished work belong here alongside practical advice and critical reflection. These choices are editorial judgments, informed by simulated audience perspectives and inclusive-design guidance, rather than measured reader preferences. They have not received comprehensive independent human review.
Broader random discovery keeps 1631 mechanically eligible sentences from 90 source-supported local articles. NLTK identifies literal spans, filtered to 12–45 words and at most 300 characters with source-markup, notice, repetition and context heuristics. Those fallible rules can miss fragments or historical invitations. Neither collection represents the whole archive.
Each eligible article has an equal chance, then one eligible excerpt is chosen within it. A longer article receives no extra article tickets. Each collection keeps its own article-ID cycle in this tab’s session; articles do not repeat until that pool is exhausted. Switching collections avoids the article just shown when alternatives exist. Reload draws again; in-app navigation keeps the sentence. No text, author names or visitor identity is saved by this feature. If session storage is unavailable, reload history is lost.
Read in context highlights the literal sentence with the full saved paragraph, source credits, links and separate rights. Credits describe the article, preserving coauthors and the distinction between article writers and quoted speakers. No quotation is rewritten; explanations sit outside its text.
Human — Ian G. Williams: set the engagement-focused criteria and requested implementation (CRediT: Conceptualization; Methodology; Project administration). AI — OpenAI Codex: assessed and selected excerpts, implemented the interface and performed recorded checks; exact model/build unknown/not recorded. GAIDeT: Code generation 0000017; Validation 0000023; Text generation 0000030; Quality assessment 0000044; Identification of limitations 0000046; Recommendations 0000047. NLTK performs saved segmentation; browser selection runs no new NLTK or MALLET analysis. Original writers retain source-text credit without implied endorsement.
The archive can document published work, teaching resources, participant experiences and critical reflection. It does not on its own establish program effectiveness, verified aggregate outcomes or causal contributions to the university mission. Those claims require additional evidence.
Fellows, alumni and authors should be able to find their work, understand how it is represented and request corrections. Students and faculty can follow practical questions to the original posts and GCDI resources.
Search the full text or filter by academic year and credited name. Each result opens an excerpt with an option to read the full saved text.
An account credit tells you who posted the article. Source-supported credits have checked evidence of who wrote it. Guest and participant posts are included. Fellowship status requires separate evidence.
All words finds whole words in any order. Exact phrase finds adjacent words in a single field or paragraph. Both ignore case, accents, and punctuation. Search covers published posts. Coverage of 423 published records and nine drafts.
An AI-assisted reading of 21 selected posts. Choose a theme to see the passages and the writers behind it. This is a situated interpretation directed by Ian G. Williams and drafted with AI assistance, not a collective statement approved by the writers.
The themes come from the essay. MALLET topics appear separately in Analysis & networks. The 21 examples cannot establish how common a theme is or whether writers agree.
Compare topics and word use across academic years, and see who is credited together.
The charts use saved analysis. Publication credits show how posts are attributed. Fellowship and collaboration require separate evidence.
How does average topic weight change among modeled local posts? The chart describes selected prose, not all blog activity or university outcomes.
Each bar averages the topic weights for modeled posts that match the filters that year. A post can have weight in several topics. Codex proposed the topic labels. Select a bar to check the posts.
Published includes curated/source-linked records. Local excludes those records; modeled also excludes welcome posts and texts too short for this recipe. Selected modeled applies the current name, year and credit filters. Different coverage can affect comparisons; it does not explain a topic change by itself.
Each bar shows the share of eligible local posts whose title or tags match the selected keyword list. A post can match several lists. Read the source to check whether a match is meaningful. These lists are separate from the MALLET topics and essay themes.
Which words recur, or occur more often in one year? Common and distinctive answer different questions; neither measures scholarly importance or statistical significance.
Explore the entire saved 301-post analyzed corpus or one academic year. This is a subset of the 423 published archive records. These word views have their own year control; the author and evidence filters used in other displays do not apply.
2026–2027 is partial. Equal-sized years give comparison space, not equal evidence. Expand a year to read every label; the tables retain all values.
Each bar shows the percentage of analyzed posts in that academic year that contain the cleaned word. Each post counts once. Small and partial years can change sharply with just one post.
Select a year to update the cloud and excerpts while keeping this word selected.
Up to five short sentences from different posts. Matches use the exact written word, ignoring case. The counts above use cleaned forms and may include variants that do not appear here.
| Word | Occurrences | Posts | Selected year vs. other years |
|---|
Counts use the archive’s existing cleaned word forms. The 2026–2027 corpus is partial. Word size is relative to the largest value in the current view.
These are year-exclusive cleaned forms: present in one observed academic year and absent from the other years of the saved 301-post corpus. This is different from distinctive vocabulary, which can appear elsewhere. A form is counted once here regardless of repetition. It is not a claim of originality, first use, importance or absence from excluded posts.
Bars use a common count scale. Striped = exclusive forms appearing in one post; solid = exclusive forms appearing in multiple posts; grey = forms also found in other years. Select a year to expand its sunburst and inspect its exclusive forms below. The comparison always shows all years; the list follows the year selector.
For every retained cleaned form, Python records the set of academic years containing it. A set of size one qualifies. All 5,731 qualifying forms are saved without a new frequency threshold. The evidence selector distinguishes single-post from multi-post forms; it does not redefine exclusivity. Counts, repeated occurrences and contributing posts are separate quantities. Years are ordered chronologically; within each year, table order is descending occurrences, then containing posts, then alphabetically. Search is a case-insensitive literal substring over these normalized forms.
The definition is fixed against all 15 saved academic years even when one year is selected. Names, code identifiers, spelling noise and preprocessing artifacts can qualify. A form used in several posts may still come from the same author. More posts and longer texts give more opportunities to observe vocabulary. 2026–2027 is partial with three analyzed posts. The companion is separate from the sunburst hierarchy because type exclusivity does not subdivide its topic weights or selected-word counts. No NLTK cleaning, MALLET fit, authorship or population change is made. Detailed method and attribution.
Select a word to inspect posts containing its saved cleaned form.
The word cloud ranks by occurrences, containing posts, or the positive percentage-point gap between one year’s post share and all other years’ post share. The last option requires two or more matching posts that year and is separate from the sunburst’s smoothed token score. Select a word to follow its yearly post shares, read short exact-form excerpts, or pick a related post. The count table gives exact denominators.
Worked example (invented counts): if “archive” occurs in 6 of 20 posts in one year and 10 of 100 posts in the other years, the cloud's distinctive gap is 30% minus 10% = 20 percentage points. The sunburst uses a different, smoothed token-rate score: wedge size is a share of selected scores, not a percentage of posts. Neither calculation is a significance test.
Choose a lens to change the question. Shared vocabulary counts a word once per containing post; its wedges divide all word–post incidences, including Other words. It does not divide distinct posts because each post contains many words. Subject mixtures sums the eight saved MALLET weights: each post contributes one unit across topics. Topic-first and year-first hierarchies preserve those same contributions. Topic labels remain interpretations.
Distinctive vocabulary compares each year's token rates against the other years, with smoothing and minimum counts. Five positive log₂ rate-ratio scores divide each year; these are score shares, not percentages of writing. The research guide records thresholds and a stricter-threshold sensitivity comparison. Adjacent-word phrases counts actual neighbouring content-word pairs from saved readable text, without joining words across removed words or punctuation. Raw frequency retains repeated token counts for comparison. Phrase and raw modes include their remaining counts as Other.
Shared, phrase, raw and distinctive modes give each eligible year equal space. Topic-mode years have space proportional to modeled posts. Year labels remain visible; smaller outer labels may be omitted, with their names and numbers always in the exact table and source panel. Select a year to expand its branches or use Show all academic years. The heatmap compares a common set of words/topics across all years using the same underlying calculations, with exact numbers in every cell. A dash means unavailable, not zero. It excludes Other and uses the union of the five displayed terms per year, or all eight topics.
The builder sums the saved NLTK-cleaned token files, joined to current post records by ID. It uses all retained tokens before MALLET’s additional vocabulary pruning. The historical recipe removes repeated boilerplate and introductory credit lines, tokenizes words, lowercases, normalizes possessives, applies part-of-speech-aware WordNet lemmatization, then removes English and documented custom stopwords and short tokens. A lemma groups forms such as plural and singular; these English-language guesses can be wrong. Read source text for context, negation and multilingual writing.
The saved manifest includes 301 of 423 published records; curated/source-linked posts, welcome/roster material, exact cleaned duplicates and items below the 30-cleaned-token threshold are excluded by the documented recipe. Drafts are excluded. Academic years use the saved August–July assignment; 2026–2027 is partial in the November 9, 2026 export. No later years or missing observations are invented. Raw counts weight longer posts more heavily; differences may reflect post volume, genre, author mix, exclusions and incomplete years. Frequency does not establish a theme, consensus, impact or fellowship membership. Word and phrase frequencies remain separate from the essay’s interpretive themes; Subject mixtures explicitly uses the saved MALLET model.
No NLTK cleaning or MALLET fitting runs in the browser. Python aggregates saved tokens, extracts literal phrase candidates and calculates versioned distinctiveness scores during the build; controls select saved results. Ian G. Williams requested these views and their explanation; OpenAI Codex implemented the aggregation, visualizations and checks. Original writers retain source-text credit, without implied endorsement. Human verification of all counts or interpretations is not established.
Saved NLP recipe · Inclusion ledger · English stopwords · Custom stopwords · Word-view methods and limitations
Follow recorded credits to the posts behind them. These links describe publication credits, not fellowship appointments or social relationships.
Choose a question to open the views and filters. You can change views at any time.
Lines count published local posts carrying the recorded credits; thicker lines mean more such posts. Name ordering reflects recorded counts, not importance. Mentoring, organizing and uncredited support are not measured by these links. Curated/source-linked records and drafts are excluded. A shared byline does not establish a working relationship or verified individual productivity. Select a name, year or line to inspect sources.
Hover over or focus a name, year, or line for a preview.
Use the count matrix to explore individual years at this width.
A shared post counts once for each pair of credited names. One group article can create many pairs. The graph records shared bylines. Collaboration requires separate evidence.
Select a year, network connection or author to inspect the source posts.
Tagging the Tower brings together writing by GC Digital Fellows and other contributors, including students who wrote about taking part in the Digital Research Institute.
Start with named writers of local posts. Use the filters to include curated works or narrow the list to source-supported credits. Select a name to see the posts and credit evidence.
Send a missing or incorrect byline to Ian G. Williams. Check unresolved posting accounts in the archive ledger.
The GC Digital Fellows help students and faculty at the CUNY Graduate Center use digital methods in research and teaching through workshops, consultations, and shared resources. The chart starts with all documented GCDI team members, including people with no credited posts. Choose a year to see the people and records behind its count.
Team scope and role also change the chart above. Scroll sideways in the table to compare years. People can appear under more than one label. A dash means annual roster evidence is unavailable.
Academic years run August–July. A directory range of 2012–2015 spans three academic years. Returning Fellows count in each year they appear and once in the overall total. The directory gives no exact appointment dates.
Article bylines identify credited writers. The GCDI People directory and annual reports help document fellowship roles and years. Archived copies of the blog helped restore 43 bylines from before 2016. The project keeps each recovered credit alongside the later WordPress posting account so readers can see the difference.
The explorer includes 423 published records, including labeled curated and source-linked entries. The saved MALLET model covers 301 posts. The interpretation uses 19 selected examples.
The reports document workshops, consultations, projects, and staffing. Eight reports provide evidence for ten academic years. Reports were not located for 2017–2018 through 2019–2020. The supplied 2025–2026 report is awaiting manual addition.
Read the report comparison and its limits · Report sources and PDF provenance (CSV) · Reported activity and PDF pages (CSV) · Reported roles and PDF pages (CSV).
Full archived articles support 43 restored pre-2016 writer credits. The later posting account is preserved as migration evidence. Report staffing, archived roster pages, and project credits describe other kinds of work and are kept separate from article bylines.
Reviewed credits with archived article links (JSON) · Archived About, People and program pages (CSV) · Archived GCDI About page with report links.
Named credits include source-supported and declared article/source-writing names in published records. The checked-only option is narrower; neither proves complete human verification. Existing reviewed identity merges are reused. Coauthored posts contribute each named writer, deduplicated across posts. Organizations and unresolved posting-account-only labels are excluded. Draft-only names are excluded.
The Digital Fellow count intersects those writers with dated GC Digital Fellow records or the human-confirmed 2026–2027 Digital Fellow roster. The card separates those bases. Broader GCDI Fellow families include Digital Fellows and other Fellow roles; they overlap, rather than add to, the Digital Fellow count. Neither is automatically membership at every post date. The summary separately reports recorded matches at publication; summer and missing evidence are not inferred.
Academic-affiliation counts use documented program/discipline labels from saved alumni claims and the undated current-team snapshot, with coverage shown. Cosmetic Earth/Environmental Science and Theatre spelling variants, one obvious Criminal Justice typo and the LAILAC abbreviation are grouped for display; raw statements remain inspectable. Broader disciplines, subfields, degree levels, joint affiliations and institutions are not merged or inferred. These labels are not a complete census of formal university programs or standardized disciplines. Annual matrices use the affiliations carried by each dated directory claim or, for the confirmed 2026–2027 roster, the saved current-team affiliation. The evidence bases remain labelled.
GCDI program/role families are a different dimension from academic programs. A person can hold several roles; missing directory matches neither establish guest status nor erase a named writer. New text models or author identities are not inferred here. Ian G. Williams requested these counts and cohort views; OpenAI Codex implemented the saved-data joins, visuals, methods and checks. Full people-count methods, source links and human/AI attribution.
Ian G. Williams supplied the project direction, audience priorities and revision requests. OpenAI Codex developed the interface, wrote code and performed checks using saved archive evidence. Ian approved this public interface. That does not establish complete human verification of the data, code or interpretations.
CRediT records documented human contributions; GAIDeT records tasks delegated to AI. Original post authors retain source-writing credit and are not automatically contributors to this new analysis.
Full project contributor disclosure ↗Search compares words and phrases against saved titles, credits/aliases, tags and complete saved plain-text reading copies. Images, embeds and original formatting are outside that search. Random selection gives eligible records equal chance and avoids the previous draw when another post is available.
Reading excerpts are shortened from the saved plain-text copy; images and embeds remain at the original source. The word and topic displays reuse existing NLTK/MALLET outputs. No new text model is trained or inferred here.
Project workflow diagrams and text versions show how the explorer was built. The analysis recipe and interface code are annotated in the private source project. Data quality, original credits, uncertainty and reviewer access are preserved as design requirements.
A targeted check of six saved posts compared readable text, cleaned tokens and the institution-related stop list. In “Doing humanidades digitales at CUNY,” the saved cleaned text retains “humanidades digitales” and “español,” while removing institution labels such as CUNY and fellow. This is a documented recipe choice, not evidence that the removed words lack meaning.
English grammar processing is not a multilingual language assessment. This small audit does not validate translation, grammatical tags or topic quality. The original reading text remains available. The separate cleaning-comparison preview compares already fitted recipes; these checks did not fit another model.
Factual corrections and interpretive disagreements are different contributions. Send a post link, the claim or excerpt you are questioning, and your evidence or alternative reading to Ian G. Williams.
Send a factual correction · Suggest an interpretation, context or excerpt change
Ian is the project maintainer and receives these requests. Suggestions are not automatically accepted; changes should retain the earlier claim, evidence and decision in the project record. No response time is promised. Source authorship does not imply endorsement of this analysis. A feedback route does not establish community approval or shared governance; broader contributor review remains to be arranged.
These revisions draw on Rawson and Muñoz on cleaning as interpretation and the Design Justice Network principles on participation and accountability. These references inform the questions; they do not certify this project.
The default appearance uses Pixelify Sans by Stefie Justprince and the Pixelify Sans Project Authors for headings, and Atkinson Hyperlegible Next by Braille Institute, Applied Design Works, Elliott Scott, Megan Eiswerth, Letters From Sweden and the Atkinson Hyperlegible Next Project Authors for reading text and controls. These designers created the typefaces; their inclusion does not imply endorsement of this project.
Both fonts are hosted locally under the SIL Open Font License 1.1, separate from the project’s MIT license. Pixelify Sans license · Atkinson Hyperlegible Next license. Historical appearances remain available in Design history.
Choose a design year to restyle every explorer view, including navigation and reading. Content, credit evidence and academic-year filters stay the same.
The designs draw on sources from 2012–2026. Each preset adapts type, shapes, and the light palette. Dark mode keeps its reading contrast. The presets interpret design trends from each year.