The workspace

A record of how the book was made: the components, the procedure, and the files. Most readers will want the draft checker or the corpus search instead. This page is for anyone who wants to see the machinery, check a number, or run the analysis on their own corpus.

Components

The corpus

103 economics research articles, tagged so that programs can see their structure — <h> around section headings, <hs> for subsections. A companion corpus of 32 Economics Letters papers is analysed separately, because the genre differs. See the corpus.

The definitions (9 files)

Word lists and patterns deciding what counts as what: hedges, boosters, connectives, signposts, move steps, reporting verbs with their stances, section-heading categories, and the language of data and estimation sections. Data, not code — they can be edited without touching a program. See methods.

The measures (13 modules)

One per unit of the book: structure, introductions, niche strategies, conclusions, abstracts, moves, section-level analyses, hedging, tense and voice, citation, sentences, connectives, structural families. Each emits statistics, verbatim examples, and notes recording its own caveats.

The library

The parser (which reads the tagged corpus, including the rule that an abstract ends at the first subsection heading), one canonical tokenizer used by every count, a part-of-speech cache for tense and voice, and the example extractor.

The manual codings

Two judgment calls no program can make: the structural family of each paper (103 coded) and the niche strategy of each introduction (98 coded, one excluded as unreadable). Coded by hand from evidence sheets the pipeline generated, with the criteria recorded alongside. Programs read the finished sheets, so even these percentages regenerate.

The outputs

results.json — every statistic, keyed; a readable digest; and print-ready tables in the book's format, placed in the layout so that a data change means a relink rather than retyping.

The quote auditor

A separate tool that reads the book's exported text and checks every quoted sentence against the corpus, sorting them into verbatim, truncated, silently edited, invented, or acknowledged (constructed illustrations, listed explicitly).

Procedure

1. Validate before trusting. Every definition file was reviewed for where its contents came from — some had been reverse-engineered from an early draft, some derived from the corpus, some were simply judgment. The ones that could not defend themselves were rebuilt.
2. Test categories against real occurrences. Concordance and collocate checks on every doubtful word. This is where bare fact left the booster list, several families became frames rather than words, and the negation problem surfaced: “no clear evidence” had been counting as boosting.
3. Clean the corpus visibly. Three extraction fragments excluded with reasons; Economics Letters split out as a separate genre; a heading pattern fixed that had been treating two articles' content as references. Every decision is recorded in the manifest rather than applied silently.
4. Replace invented numbers with measured ones. The connective “target ranges” had been set by hand; they became quartiles computed from the corpus.
5. Code the judgment calls explicitly. Batch by batch, with the evidence visible — headings, subheadings, theorem apparatus, candidate sentences and their position in the introduction. The conventions that emerged are written down.
6. Reconcile the book against the pipeline. Every printed claim matched to its source, with the reason for any difference. Several headline findings changed.
7. Audit claims, not just numbers. Wording was checked against what the measurement actually supports; several “the corpus suggests…” statements became plainly labelled advice.
8. Certify the result. The exported book goes through the quote auditor and a number sweep; the report says what remains. It runs in seconds and can be repeated on any future edition.

Some things that changed

The reconciliation was not cosmetic. Among the findings that moved:

• The leading niche strategy is continuation (35%), not gap-filling (28%). The earlier figure had let the frequency of however stand in for gap-creation.
• Genuinely thematic architecture is rare (3% of papers, not 24%): most apparently thematic papers keep a functional skeleton one heading level down.
• The introduction takes about a third of the paper, not 38% — the higher figure was inflated by short Economics Letters papers in the same corpus.
• The hedge-to-boost ratio inverted once negation was handled correctly. The finding that survived is the one that mattered: hedging rises from results to conclusion.

Files

The data this site runs on is the analysis output itself:

• articles.json — metadata and codings for all 103 articles
• corpus.json — the section text used for concordancing
• index.json — 40,569 pre-computed category hits
• definitions.json — the word lists, as the analysis reads them
• norms.json — the comparison distributions used by the draft checker

Downloads

• econ-corpus.zip (1.5 MB) — the tagged corpus of 103 articles, with the full attribution and licence file
• corpusbook-code.zip (110 KB) — the complete analysis pipeline: parser, measures, definitions, coding sheets, criteria, and instructions for running it
• ATTRIBUTION.txt — citations for all 103 articles

The corpus text is redistributed under CC BY 4.0; the code may be used freely with attribution. Running the pipeline on your own corpus needs only a manifest file and corpora in the same tagged format — see HOW_TO_RUN.txt in the code archive.