Methods you can inspect.

Reproducible calculations

NFC-normalized lowercase Unicode letter/number tokens retain internal apostrophes and hyphens. TTR = types/tokens. Guiraud = types/√tokens. Herdan = log(types)/log(tokens), undefined at one token. Shannon entropy uses base 2. MATTR averages every overlapping window; short text uses its actual token count. English sentence boundaries use Intl.Segmenter and may differ across browser Unicode versions.

English models

compromise 14.14.4 supplies contextual tag and base-form predictions. It is not gold-standard morphological segmentation or a dependency parser. CoNLL-U mode reads your independently supplied annotations and checks basic structure only: ten columns, ordered IDs, UPOS, heads, one root, root agreement and cycles. Enhanced DEPS, all UD rules and feature inventories are not validated.

Sentiment

VADER-JS 1.1.3 applies a lexicon and rules. Compound ≥0.05: positive; ≤−0.05: negative; otherwise neutral. Components are not probabilities. Fourteen upstream demo fixtures are regression-tested. Domain-specific accuracy is unverified; irony, old spelling and implicit meaning can be missed.

Register and forensic interpretation

Register Profile counts auditable surface forms per 1,000 tokens. It does not implement MAT’s 67 features, reference z-scores or Biber dimensions. Forensic comparisons use vocabulary/word-trigram set Jaccard, word-count/character-trigram cosine, shared five-word sequences and function-word/punctuation counts. Character trigrams use collapsed whitespace and UTF-16 code units. Similarity cannot establish authorship, plagiarism, deception or legal facts.

Resource budgets

Lexical/register desks: 2 million characters. English morphology/syntax: 200,000. Annotated CoNLL-U: 1 million. Sentiment: 12,000. Forensic: 500,000 per text. Limits are explicit; no silent truncation. Workers can be stopped; analysis times out at 90 seconds without claiming partial success. Tables show 500 rows; CSV exports the full selected table; JSON includes all result tables and notes. Formula-like CSV cells are apostrophe-prefixed.

Text-only conversion

Selectable PDF and DOCX text is extracted locally. Text-only Word keeps paragraphs; PDF reflows text onto A4 pages using pdf-lib, fontkit and the locally embedded Liberation Sans font. Images, tables, columns and original pagination are not reconstructed. NFC and line endings are normalized; tabs become four spaces; whitespace reflows; long unbroken strings may wrap. Unsupported glyphs fail export without discarding content. PDF output supports 2 million characters and 2,000 pages per document, subject to memory. Scanned pages need OCR first; old .doc is unsupported.

Privacy and sources

Inputs and analysis stay in the browser. No AI API, text telemetry, account or server upload is used by these desks. Source and runtime versions: compromise, VADER-JS, CoNLL-U specification, pinned bundles.

compromise license · VADER license · VADER notice · fontkit license · pako license