# StyleBot sources and exclusions

Research snapshot: 9 September 2026. These are provisional references for discussing writing choices, not ideal sentence ratios or grading targets.

[Read the six-page study](/stylebot/benchmark-study.pdf), [view the discipline charts](/stylebot/discipline-benchmarks.png), or [download the complete source metadata](/stylebot/source-audit-summary.json).

## What the collection contains

| Collection | Source accounts | Eligible works | Prepared main-prose words |
|---|---:|---:|---:|
| Supplied PDFs | 101, including one excluded PDF | 102 | 721,582 |
| Added scholarly sources | 44: 13 PDFs and 31 HTML articles | 44 | 441,885 |
| Combined | 145 accounts; 144 eligible sources | 146 | 1,163,467 |

A source file and an authored work are different units. One supplied PDF contains three separately credited contributions. The complete metadata lists every original PDF once and nests those contributions beneath their shared source. Each included external article appears once. Names, broad writing fields, page ranges, prepared-word counts, exclusion reasons, available publisher URLs and source-file hashes are retained. Source essays and private annotations are not redistributed.

| Writing field | Original works | Added works | Total works |
|---|---:|---:|---:|
| Literary criticism | 39 | 10 | 49 |
| History | 24 | 3 | 27 |
| Social research | 22 | 11 | 33 |
| Legal scholarship | 17 | 20 | 37 |

These are purposively collected scholarly texts, not a random national sample. Broad fields can include work across disciplines. Author names follow source attribution; faculty rank has not been independently verified for every author.

## Significant inclusion decisions

- **S047 — Debating Patriarchy:** Julia Adams’s *Patriarchy Resartus*, Benita Roth’s *Patriarchy and Intersectionality: The Price of Popularity and the Project of Recovering Meaning*, and Pavla Miller’s *Where Next?* are three works in one PDF. Their prose occupies PDF pages 2–8, 8–12 and 12–16. Introductory framing was excluded because its individual authorship was not separately established.
- **S058 — Rural Puerto Rico in the Early Twentieth Century Reconsidered: Land and Society, 1899–1915:** the work by César J. Ayala and Laird W. Bergad was excluded. Right-edge clipping was visually confirmed on PDF page 30 in the supplied file and the official publisher replacement. Reliable whole-work coverage could not be established; this does not mean every page was independently found damaged.
- **S070 supplementary Word file:** *The Color of Money and the Nature of Value: Greenbacks and Gold in Postbellum America*, by Bruce G. Carruthers and Sarah Babb, is already represented by its PDF. Matching publication metadata and extractable text identify the Word file as the same work. Its 719 extractable word tokens and three embedded images were not established as a fully transcribed or byte-identical duplicate. It adds no independent essay.
- **L018 — The Democratic Deficit of Living Originalism:** the official repository landing page was obtained, but complete article text was unavailable. It adds no words. L025 is a different accessible work by the same author, Stephanie Hall Barclay.
- **Whole-issue and unrelated downloads:** three Mediations PDFs obtained through article navigation contained entire journal issues; only the individually attributed article HTML for E006, E007 and E008 was used. An unrelated Harvard Law Review factsheet link was initially selected for two other candidates and replaced before preparation. Its rejected binary and immutable receipt were not retained, so the metadata identifies the historical URL without assigning it a replacement article’s hash.

## What was removed during preparation

The preparation considered usable main prose throughout each work, rather than selecting convenient excerpts. Rules and source checks addressed front matter, titles and bylines, running headers and footers, notes, references, tables, captions, displayed quotations, reproduced verse, transcripts and damaged extraction. Inline quotations within an author’s argument generally remain.

| Preparation stage | Counted unit | Exclusion decisions |
|---|---|---:|
| Original PDF extraction | Extracted PDF line | 38,960 |
| Original later exclusions | Prepared paragraph | 136 |
| External PDF extraction | Extracted PDF line | 10,897 |
| External HTML extraction | Paragraph inside the inspected article body | 769 |
| External later exclusions | Prepared paragraph | 2 |

**These counts cannot be added into one total of removed words or sentences.** Lines and paragraphs are different units, and later decisions can supersede earlier ones. The 136 original-paragraph exclusions comprise 116 paragraphs covered by the S058 whole-work exclusion, 15 source-gap paragraphs, four independent quotation candidates, and one truncated ending. The two later external-paragraph exclusions are epigraphs in L017. S047’s later authorship decision is separate from its initial parent-PDF line counts; overlapping contribution-specific line ledgers are not summed.

PDF page numbers refer to pages in the file, not printed journal numbering. A reported page range is an envelope of retained prose and may contain interior omissions. HTML sources have article-body paragraph locations instead of PDF page numbers. A paragraph without terminal punctuation is a boundary-review flag, not automatically an excluded paragraph or proof of source damage.

## Corrections after the original freeze

A later source review confirmed **27 correction operations across 22 works**. They address leaked headings and captions, note markers, word extraction, split or merged sentence boundaries, page continuations, omitted main prose and a reproduced ballad block. The separate correction scenario replaces 90 frozen units with 39 units, changing 40,875 candidate sentence units to **40,824**, while retaining all 146 work identities.

The published reference uses this source-corrected scenario. The independent audit figures in the report remain tied to the original frozen text; the corrected corpus has not received a new untouched validation sample. Sixteen other flagged candidates were outside this bounded source-confirmed correction ledger. The complete metadata keeps initial preparation and later correction operations separate.

Preparation, source checks and AI review do not establish perfect transcription or human-validated grammar accuracy. The resulting percentages are descriptive estimates. Students and PLFs can use them to discuss sentence choices and their effects; matching a pie chart does not demonstrate learning.

The PDF remains a dated research snapshot. Its statement that the research pass had not changed a production deployment describes that earlier research stage.