A lightweight, privacy-focused web application designed for students and professors at Utica University to analyze academic text for stylistic patterns, readability metrics, syntactic structure, and deterministic markers of machine-generated text.
-
100% Client-Side Processing: All text extractions and calculations run entirely inside your browser. No documents, text, or student data are ever uploaded to a server or stored externally.
-
Deterministic Analysis (No AI Models): Uses statistical algorithms, mathematical formulas (such as Shannon Character Entropy, Moving-Average Type-Token Ratio, and Zipf's Law distribution fitting), and explicit pattern matching rather than black-box AI detection models.
-
Single Document & Text Analysis: Evaluate pasted text or individual
.docx,.pdf, and.txtfiles with immediate breakdowns of tone, readability, structural symmetry, and stylometrics. -
Moodle Batch ZIP Processing: Native support for Moodle's "Download all submissions"
.zipfiles. Automatically maps student directory names to their attached documents and parses each file. -
Batch Score Matrix & Histogram: Generates an interactive submission matrix alongside a visual distribution histogram (powered by Chart.js) categorizing submissions from Highly Likely Human to Highly Likely AI / Bot.
| Category | Metric | Description & Target Thresholds |
|---|---|---|
| Authenticity & Stylometrics | Human Authenticity Score | Overall composite score based on sentence length variance, lexical richness, and artifact penalties. Expected: >55%. |
| Authenticity & Stylometrics | Shannon Character Entropy | Measures string distribution predictability based on character frequency. Expected: >4.0. |
| Authenticity & Stylometrics | Lexical Diversity (TTR) | Ratio of unique words to total words in the passage. Expected: >0.45. |
| Authenticity & Stylometrics | Sentence Variance | Measures rhythm and variance in sentence length to detect monotonous syntax. Expected: >25.0. |
| Authenticity & Stylometrics | Artifact & Flag Checks | Detects synthetic citations (e.g., arXiv/DOI hallucination patterns), repetitive sentence openers, rigid LLM section headers, and AI buzzword overuse. |
| Readability | Flesch Reading Ease | Score from 0–100 measuring structural reading accessibility. Technical/academic text typically scores below 40.0. |
| Readability | Flesch-Kincaid Grade Level | Translates sentence length and syllable density directly into U.S. educational grade levels. |
| Prose Structure | Nominalization Ratio | Percentage of abstract "zombie nouns" converted from verbs (words ending in -tion, -ment, -ance, -ence). Ratios above 5.0% indicate bloated prose. |
| Prose Structure | Sentence Clause Complexity | Percentage of multi-clause or conjunction-heavy sentences. Expected: >30% for complex academic prose. |
| Tone & Punctuation | Hedging vs. Assertive Terms | Counts cautious terms (suggests, perhaps, might) against assertive terms (definitely, proven, clearly). |
| Tone & Punctuation | Punctuation Fingerprint | Tracks usage frequency of em-dashes (—), semicolons (;), exclamation marks (!), and questions (?). |
-
Open
index.htmlin any standard web browser. -
Paste text into Option 1, or select a
.txt,.docx, or.pdffile under Option 2. -
Click Run Comprehensive Analysis to generate the detailed metric card.
-
In Moodle, select Download all submissions for an assignment to receive a
.ziparchive. -
Select the
.ziparchive under Option 2 and click Run Comprehensive Analysis. -
The tool will parse each subdirectory, run individual stylometric calculations, and display:
-
A Score Distribution Histogram summarizing author risk categories.
-
A Submission Score Matrix Table listing student name, document title, authenticity score, verdict classification, MATTR, Zipf
$R^2$ , passive voice percentage, and triggered flags.
-
PowerPoint Files (
.pptx): Do not upload PowerPoint presentations. Due to fragment sentence structures, bulleted layouts, and minimal overall text length, presentation slides will produce inaccurate statistical readings. -
Thesis & Dissertation Documents: Long-form graduate work features highly specialized terminology, technical iteration, and formal cadence that can occasionally skew baseline metrics. Scores should be evaluated manually alongside context.
-
Interpretative Guidelines: Results are meant to serve as structural flags for direct pedagogical review, not absolute proof of non-human authorship.
This application is built as a single-page HTML5/JS application relying on client-side Web APIs and CDN-hosted browser libraries:
-
PDF.js: Client-side PDF text content extraction.
-
Mammoth.js: Client-side
.docxraw text extraction. -
JSZip: In-memory unzipping and directory traversal of archive files.
-
Chart.js: Canvas-based visual histogram rendering for batch analytics.
-
GitHub Repository: https://github.com/securemindorg/Utica-Document-Analysis
-
Organization: SecureMind / Utica University Stylometrics Project