TURNITOUT
# Intelligent Similarity Evasion & Multi-Algorithm Checker

Turnitout User Manual & Technical Reference

OFFICIAL DOCUMENTATION

Welcome to the authoritative user manual for Turnitout Core. This technical guide is designed for power users, researchers, and developers who want an exhaustive reference for every module, transformation knob, slider, document engine, and formatting safety guarantee in the application.

"Turnitout operates as a strictly deterministic, zero-generative-AI linguistic transformation engine. Everything described in this manual operates in transient serverless RAM with zero disk persistence and zero content logging."

1. System Architecture & Deterministic Philosophy

Conventional commercial AI rewriters (QuillBot, Undetectable AI, StealthGPT) route text through third-party Large Language Models (LLMs). This introduces three fatal flaws for academic and professional documents:

  1. Factual Hallucinations: Generative models randomly invent statistics, dates, citation numbers, and theorem claims.
  2. Secondary AI Fingerprints: LLM output inherently exhibits low mathematical perplexity and uniform sentence length, frequently triggering detectors like Turnitin AI and CopyLeaks anew.
  3. Formatting Destruction: Plaintext LLM wrappers obliterate complex Word XML styles, LaTeX math environments, tables, equations, and code indentation.

The Turnitout Paradigm (Black-Box Guarantee): Turnitout uses zero generative AI. It functions as a deterministic linguistic transformation pipeline. Input text passes through structured syntactic parsing, contextual synonym realignment, n-gram chain disruption, and discourse connective rotation. All facts, equations, numerical values, and document layouts remain mathematically sealed.

2. Core Module: Plagiarism & AI Humanizer

The primary humanization workspace provides dual operating modes: direct prose entry and multi-format document ingestion.

2.1 In-Place Document Parsing & Layout Preservation
Word (.docx) Excel (.xlsx / .csv) LaTeX (.tex) Markdown (.md) Plain Text (.txt) Code (.py) Batch ZIP Archives

Unlike converters that flatten documents to plain text, Turnitout operates directly on document Abstract Syntax Trees and XML run streams:

  • Microsoft Word (.docx): Traverses w:p (paragraphs) and modifies only inner w:t text runs. Formatting runs (bold, italic, font face, text highlighting, paragraph alignment, headers, footers, tables, and drawing objects) remain untouched.
  • Excel Spreadsheets (.xlsx / .csv): Iterates through tabular text cells and comments, humanizing prose while preserving cell formulas, numbers, dates, and column structures intact.
  • LaTeX (.tex): Isolates natural language body prose while mathematically protecting math modes ($...$, \[...\]), environments (\begin{equation}, align, matrix), citation anchors (\cite{...}), and document preambles.
  • Source Code (.py, etc.): Extracts and humanizes only natural language docstrings and comments, keeping all executable code statements, variables, and AST logic 100% operational.
  • Batch ZIP Processing: Upload archives containing multiple documents or codebases. Turnitout processes each file in-memory and delivers a bundled ZIP export (Turnitout_Humanized_Docs_<TIMESTAMP>.zip) preserving relative path hierarchies.
2.2 Watermark, Hidden Unicode & Metadata Scrubbing
Anti-Tracking EU AI Act Compliance Automatic

Turnitout automatically scrubs modern tracking footprints embedded by generative AI providers to comply with watermarking regulations:

Tracking Mechanism How Providers Inject It Turnitout Functional Outcome
Zero-Width Unicode Spaces Hidden characters (U+200B, U+200C, U+200D, U+2060) inserted between words as steganographic watermarks. Deep regex character scrubbing detects and purges all invisible zero-width code points without altering visible spacing.
C2PA Manifest Metadata Provenance manifests embedded in Office and PDF document packages indicating AI generation origins. Document packaging streams are sanitized upon export, wiping internal generation markers from file history.
LLM Cliché Signatures High-frequency AI marker phrases (e.g. "delve into", "pivotal role", "testament to", "tapestry") and em-dashes (—). Targeted catalog disruption identifies cliché n-grams and substitutes natural, varied academic phrasing.
2.3 Advanced Transformation Controls (The 14 Knobs)
Fine-Tuning Interactive Controls Live Output Previews

Inside the expandable Advanced Transformation Settings & Protected Terms panel, users can customize individual syntactic passes. Below is an exhaustive breakdown of every knob, including its functional role, word count impact, and live before/after output comparisons.

1. Synonym Swapping (with Alignment Probability Slider)

Word Count Impact: 0 Words (1:1)

What It Changes: Targets statistically predictable single words overrepresented in AI text and replaces them with contextually accurate, lower-frequency synonyms without adding extra tokens. The Synonym Alignment Probability Slider (0.50 to 1.00) regulates semantic precision: 0.90 – 0.95 guarantees strict formal academic fidelity with zero semantic drift; 0.80 – 0.85 offers optimal natural flow; checking Randomize Alignment simulates human drafting variance.

Original AI Input
"The algorithm demonstrates substantial efficacy in mitigating network latency across distributed clusters."
Transformed Output (1:1 Swap)
"The algorithm shows significant effectiveness in reducing network delay across distributed clusters."

Recommendation: Always keep ON. This is the primary foundation for breaking AI n-grams while strictly maintaining 100% word-count neutrality.

2. Phrase Rewriting

Word Count Impact: -2 to +1 Words

What It Changes: Scans for hallmark multi-word AI clichés, formulaic transitional phrases, and boilerplate expressions, replacing them with direct, authentic human academic phrasing.

Original AI Input
"In order to delve into the multifaceted dimensions of this phenomenon, it is of paramount importance to consider..."
Transformed Output
"To examine the various aspects of this phenomenon, one must consider..."

Recommendation: Highly recommended ON. Strips recognizable AI patterns like "delve into", "multifaceted", and "paramount importance".

3. Determiner Swapping

Word Count Impact: 0 Words (1:1)

What It Changes: Contextually alternates articles ("the", "a") and demonstratives ("this", "these", "such"). AI models frequently start almost every sentence with "The [subject]"; this pass creates organic grammatical diversity.

Original AI Input
"The experiment validates the hypothesis. The results suggest the model generalizes well across the unseen test datasets."
Transformed Output
"This experiment validates the hypothesis. These results suggest the model generalizes well across such unseen test datasets."

Recommendation: Keep ON. Eliminates repetitive sentence openings with exact 1-to-1 word count preservation.

4. Voice Transformations (Active ↔ Passive)

Word Count Impact: ± 1 to 3 Words

What It Changes: Converts rigid passive constructions into dynamic active voice, or converts repetitive active statements into scholarly passive constructions where objective agency is preferred.

Original AI Input (Rigid Passive)
"A novel optimization framework was engineered by the research group to accelerate gradient descent."
Transformed Output (Dynamic Active)
"The research group engineered a novel optimization framework to accelerate gradient descent."

Recommendation: Keep ON for standard humanization. Elevates stylistic burstiness.

5. Sentence Fusion

Word Count Impact: +1 to 3 Words

What It Changes: Merges adjacent short, staccato sentences into cohesive compound or complex sentences using appropriate coordinating or subordinating conjunctions.

Original AI Input (Choppy Staccato)
"The dataset contains 50,000 samples. The training process required 12 hours. The final validation loss was minimal."
Transformed Output (Fused Compound)
"The dataset contains 50,000 samples, and while training required 12 hours, the final validation loss remained minimal."

Recommendation: Enable for fragmented AI drafts. Disable if targeting a strict exact word count limit.

6. Transition Injection

Word Count Impact: +2 to 4 Words / Sentence

What It Changes: Injects authentic scholarly discourse markers and transitional signposts (e.g., "Furthermore", "In contrast", "Consequently", "To this end") to enhance flow.

Original AI Input
"The baseline model struggled with long-range dependencies. The attention-augmented architecture maintained high accuracy."
Transformed Output
"The baseline model struggled with long-range dependencies. In contrast, the attention-augmented architecture maintained high accuracy."

Recommendation: Excellent for essays and literature reviews. DISABLE if you have a rigid word limit ceiling.

7. Clause & Word Reorder

Word Count Impact: 0 Words (1:1)

What It Changes: Reorders dependent clauses, adverbials, and prepositional adjuncts to natural alternative positions without altering semantic meaning or document length.

Original AI Input
"The model achieves superior convergence speeds when evaluated on GPU clusters."
Transformed Output
"When evaluated on GPU clusters, the model achieves superior convergence speeds."

Recommendation: Always keep ON. Radically disrupts predictable sequence starts with zero word count inflation.

8. Nominalization / Verbalization

Word Count Impact: ± 1 to 2 Words

What It Changes: Alternates between direct action verbs and abstract nominal forms (e.g., "analyze" ↔ "conduct an analysis of", "measure" ↔ "perform a measurement of").

Original AI Input
"We systematically measured the latency variances across all worker nodes."
Transformed Output
"We conducted a systematic measurement of latency variances across all worker nodes."

Recommendation: Keep ON for formal scientific manuscripts and academic theses.

9. Appositive Insertion

Word Count Impact: +5 to 15 Words / Instance

What It Changes: Inserts clarifying parentheticals, informative noun phrases, or acronym definitions directly alongside subjects, replicating authentic human explanatory prose.

Original AI Input
"We evaluated BERT on the SQuAD benchmark."
Transformed Output
"We evaluated BERT, a transformer-based language representation model, on the SQuAD benchmark."

Recommendation: Excellent for deep explanatory papers. DISABLE if you are bound by a strict word limit ceiling.

10. Discourse Rotation

Word Count Impact: 0 to +2 Words

What It Changes: Rotates macro-connectives and paragraph openers according to discourse tree taxonomies, preventing adjacent paragraphs from opening with identical rhetorical words.

Original AI Input (Repetitive Connectors)
Paragraph 1: "Moreover, the system minimizes query overhead."
Paragraph 2: "Moreover, the cache structure improves latency."
Transformed Output (Rotated Diversity)
Paragraph 1: "Moreover, the system minimizes query overhead."
Paragraph 2: "Concurrently, the cache structure improves latency."

Recommendation: Essential for multi-page documents, research reports, and dissertations.

11. Contractions Swapping

Word Count Impact: +1 Word / Expansion

What It Changes: In academic presets, expands informal contractions into rigorous formal pairs (e.g., "don't" ➔ "do not", "can't" ➔ "cannot"). In casual mode, contracts them for natural dialogue.

Original AI Input (Informal Contractions)
"It's evident that the pipeline didn't encounter memory fragmentation."
Transformed Output (Academic Standard)
"It is evident that the pipeline did not encounter memory fragmentation."

Recommendation: Turn on for formal academic submissions.

12. N-Gram Chain Breaking

Word Count Impact: 0 to ±1 Word

What It Changes: Audits surviving 5-gram token chains against precomputed statistical AI sequences and applies targeted micro-adjustments to shatter hash fingerprints checked by Turnitin and GPTZero.

Original AI Input (Predictable 8-Gram Chain)
"...serves as a testament to the remarkable capabilities of modern deep learning..."
Transformed Output (Shattered Sequence)
"...highlights the notable performance and capabilities of modern deep learning..."

Recommendation: Always keep ON. Critical defense against Winnowing and hash fingerprint algorithms.

13. Conceptual Bridging

Word Count Impact: +10 to 25 Words / Paragraph

What It Changes: Analyzes conceptual relationships between adjacent paragraphs and generates an introductory thematic bridge to smoothly link disparate arguments.

Original AI Input (Abrupt Transition)
[End of Para 1]: "...concluding our evaluation of loss functions."
[Start of Para 2]: "Distributed training requires specialized hardware."
Transformed Output (Smooth Thematic Bridge)
[Start of Para 2]: "Building upon these loss function dynamics, distributed training necessitates specialized hardware..."

Recommendation: Use for capstone theses and research manuscripts requiring high thematic narrative cohesion. DISABLE for strict word limits.

14. Information Reordering

Word Count Impact: 0 Words (1:1)

What It Changes: Reorders sentence clauses according to the linguistic "given-before-new" principle, placing foundational context at the start and reserving the climax for key conclusions.

Original AI Input
"Substantial throughput gains emerged when the engineers parallelized the pipeline."
Transformed Output
"When the engineers parallelized the pipeline, substantial throughput gains emerged."

Recommendation: Keep ON. Optimizes reader emphasis and sentence cadence with zero word count inflation.

2.4 Custom Protected Terms (Synonym Exclusion Blacklist)

Certain technical terminology, trademarked product names, chemical formulas, or variable names must never be substituted. The Protected Terms system ensures these strings remain 100% untouched.

Method A: Manual Text Entry

Type comma-separated terms into the text field:
CRISPR-Cas9, TensorFlow, O(N log N), React

Method B: JSON File Upload

Upload a .json list or object. Valid schemas:
["term1", "term2"] or {"words": [...]}

Original AI Input
"We initialized the AdamW optimizer with a learning rate of 0.001 to train the BioBERT architecture on our custom clinical dataset."
Transformed Output (Protected: AdamW, BioBERT, learning rate)
"We configured the AdamW optimizer with a learning rate of 0.001 to fine-tune the BioBERT architecture on our custom clinical dataset."
2.5 Empirical Voice Matching (Beta)

When writing high-stakes submissions, tone consistency is vital. The Empirical Voice Matching engine calibrates transformation parameters to match an author's historical writing sample:

  • Upload an authentic past essay, published paper, or code documentation sample (100+ words).
  • The engine analyzes stylistic heuristics: sentence length distribution, clause depth, contraction preferences, and lexical formality.
  • Transformation probabilities are dynamically calibrated so the resulting output matches your authentic authorial cadence.
Generic AI Output (Without Voice Calibration)
"Furthermore, it is of paramount importance to recognize that the dataset presents multifaceted challenges regarding class imbalance."
Voice-Calibrated Output (Mimicking Direct Author Style)
"Crucially, the dataset presents substantial challenges regarding class imbalance."
2.6 Telemetry & Transformation Breakdown

Upon completing humanization, open the Telemetry & Transformation Breakdown expander to inspect performance metrics:

  • Word Count Delta: Exact before & after word count comparison to ensure compliance with length boundaries.
  • Transformation Intensity: Total number of lexical replacements, clause reorderings, and discourse shifts applied.
  • Interactive Diff Viewer: Color-coded side-by-side comparison highlighting all modified spans for rapid verification.

3. Core Module: Code Comment & Docstring Humanizer

Designed for computer science students, software engineers, and researchers submitting programming assignments.

3.1 AST String-Aware Comment Isolation (100% Executable Safety)

The single most dangerous failure of generic rewriters on source code is altering syntax, variable names, or string literals. Turnitout features a string-aware tokenizer that tracks quote contexts (single quotes, double quotes, escape sequences) to isolate only natural language comment tokens:

# ONLY comments like this are rewritten def calculate_loss(y_true, y_pred): "String literals 'like this' are strictly untouched" return np.mean((y_true - y_pred) ** 2)

Supported languages: Python, R, JavaScript, TypeScript, Java, C, C++, C#, Go, Rust, PHP, Swift, Kotlin, Bash/Shell, SQL, and HTML/CSS/XML.

3.2 Developer Skill Personas

Match comment vocabulary to your intended developer seniority level:

  • Beginner Developer: Uses simple, everyday conversational terminology (e.g., "checks if list is empty", "loops through numbers").
  • Intermediate Developer: Standard industry working English (e.g., "validates container boundary", "iterates across dataset").
  • Senior Architect / Expert: High-density technical jargon (e.g., "evaluates invariant constraint", "executes sequential traversal").
3.3 Docstring Conversion

Generative AI tools (ChatGPT, GitHub Copilot) frequently generate verbose triple-quoted docstrings (""" ... """). Enabling Convert AI Docstrings → Human '#' Comments automatically refactors multi-line docstring blocks into natural, concise line-by-line developer comments.

4. Core Module: Citation Inserter & BibTeX Generator

Academic papers require verifiable citations. The Citation Inserter provides contextual reference insertion and automated bibliography package generation.

4.1 Citation Formats & Contextual Placement

Turnitout detects high-salience claim sentences in your manuscript and anchors citations where scholarly references are standard.

  • IEEE Format: Sequential numeric anchors (e.g., [1], [2], [3]).
  • APA Format: Author-year markers (e.g., (Vaswani et al., 2017)).
  • LaTeX Format: Exact citation macros (e.g., \cite{vaswani2017attention}).
4.2 Five Academic Taxonomies

Select your academic discipline or use Auto-Detect Field:

Computer Science & AI Engineering & Physics Mathematics & Statistics Quantitative Finance Medical & Life Sciences
4.3 Self-Executing AI Auto-Fill BibTeX Schema

When you download the generated citations.bib file, it includes an integrated prompt header. Instead of searching Google Scholar manually for DOIs:

  1. Click Download BibTeX (.bib) or copy its content.
  2. Paste the raw text directly into ChatGPT, Claude, or Gemini.
  3. The embedded header instructs the AI to fill each citation key with real, peer-reviewed papers, authors, and DOIs matching your field while strictly retaining the exact citation keys in your paper.

5. Core Module: AI Pattern & Signature Checker

Modeled after QuillBot and ZeroGPT diagnostic dashboards, this tool allows local verification before submission:

  • ZipPy Entropic Perplexity: Analyzes compression entropy ratios across contiguous text chunks. AI text compresses unnaturally well due to repetitive token probabilities; human drafting produces high entropy.
  • Composite Risk Percentage: Overall probability gauge (0% to 100%) showing overall detection likelihood.
  • Trilateral Breakdown: Separates text into AI-Generated, AI-Refined, and Human-Written percentages.
  • Interactive Highlighting: Low-entropy spans are highlighted in amber/yellow for immediate manual review.

6. Core Module: Pairwise Document Similarity

Designed for instructors, researchers, and students verifying version overlap between two separate manuscripts (Document A vs Document B):

  • Multi-Pass Cross-Audit: Simultaneously computes 5-gram overlap, Winnowing fingerprint similarity, Cosine TF-IDF, and Longest Common Subsequence (LCS).
  • Plagiarism Risk Index: Instant categorization into Low, Moderate, or High Risk overlap.
  • Side-by-Side Span Matching: Displays matching passages in synchronized scroll windows.

7. High-Availability Cluster Protocol & Privacy

To prevent server resource exhaustion and bypass Streamlit Cloud rate limits during peak academic exam seasons, Turnitout operates across an 8-node distributed cluster:

Cluster Node Live URL Operational Purpose
Server Node 1–7 turnitout1.streamlit.app ... turnitout7.streamlit.app High-speed dedicated evasion worker mirrors handling ~90% of global traffic.
Main Server (Heart) turnitout.streamlit.app Resilient cluster anchor, OAuth gateway, and fallback node.
100% In-Memory Privacy Guarantee: All document parsing, AST traversal, and text humanization execute in transient server RAM. Files are never written to physical disk, never logged to relational databases, and never retained for model training. Once your session ends, all document data dissolves.

8. Power-User FAQ & Quick Troubleshooting

Q: My transformed document grew in word count. How do I lock it to 0 added words?

Open Advanced Transformation Settings & Protected Terms. Keep Synonym Swapping enabled (slider at 0.85 – 0.90), and turn OFF Transition Injection, Appositive Insertion, and Conceptual Bridging. The engine will perform strict 1:1 substitutions with 0 net words added. Read the dedicated Word Limit Guide for full details.

Q: Does Turnitout modify formulas, citations, or figure captions in LaTeX?

No. Turnitout's LaTeX parser strictly protects all math environments ($...$, \[...\], equation, align), macros (\cite{...}, \ref{...}), tabular columns, and preambles. Only natural prose paragraphs receive linguistic realignment.

Q: How do I report a bug or request a custom academic discipline taxonomy?

Use the in-app Feature Request & Bug Reporter in the sidebar, chat directly with developer Ahmad Hassan on WhatsApp at +92 322 5522383, or submit a GitHub Issue.

Ready to Experience the Engine?

Test your documents and code with complete privacy across all 8 live cluster nodes.

Launch Live Application