# AI Detector by Ceptile — Complete Knowledge Base & Technical Documentation

> Hostname: https://aidetector.ceptile.com
> Parent Platform: Ceptile (https://ceptile.com)
> Founder & Chief Architect: Shaurya Agarwal
> Version: 2026.2 (Client-Side Forensic Engine)
> Format: Unified Single-Document LLM Context Dump

---

## 1. Executive Summary & Architecture

AI Detector by Ceptile is a multi-modal, zero-upload AI content detection and forensic analysis web application. It runs 100% locally inside the visitor's browser tab using Web Workers, WebAssembly (Wasm), and the Web Audio API. 

Unlike traditional cloud-based AI detectors that ingest student essays, corporate memos, images, or audio onto remote servers—risking intellectual property leaks, FERPA violations, and data training ingestion—AI Detector by Ceptile transmits zero bytes of user data across the network. All tokenization, statistical parsing, frequency transforms, and machine learning inferences execute deterministically on-device.

### Core Architectural Guarantees:
1. **Zero-Upload Execution:** User text, uploaded PDFs, DOCX documents, images, video files, and audio tracks never leave the browser. There is no backend analysis API endpoint.
2. **Mathematical Explainability:** The system rejects single-number black-box opacity. Every score is derived from transparent statistical formulas, showing exact sentence-level Z-scores, Zipf regressions, and feature weights.
3. **Multi-Modal Forensics:** Supports text (prose, academic papers, source code), images (photos, generative renders), video (deepfakes, generative motion clips), and audio (synthetic speech, neural vocoder music).
4. **Mandatory Human-in-the-Loop Oversight:** By contract and design, detector outputs represent probabilistic evidence, not indisputable proof. Sole-basis adverse decisions (e.g., student expulsion, academic misconduct penalties, employee termination) are strictly prohibited.

### Dedicated Canonical Multi-Page Application Endpoints:
- `https://aidetector.ceptile.com/` — Core multi-modal AI detector hub.
- `https://aidetector.ceptile.com/text-ai-detector` — Specialized text AI detector evaluating 46 stylometric signals across ChatGPT, Claude, Gemini, and DeepSeek text.
- `https://aidetector.ceptile.com/ai-checker` — Free online AI checker portal for rapid, unlimited content audits across all formats.
- `https://aidetector.ceptile.com/human-checker` — Dedicated human writing verification hub helping authors and students verify authentic human prose and defend against false accusations.
- `https://aidetector.ceptile.com/image-ai-detector` — Visual forensics inspecting C2PA credentials, EXIF headers, and Bayer CFA sensor noise.
- `https://aidetector.ceptile.com/video-ai-detector` — Temporal grain and motion optical flow dynamics for Sora, Veo, Kling, and Runway footage.
- `https://aidetector.ceptile.com/audio-ai-detector` — Spectral centroid and vocoder phase coupling analysis for ElevenLabs, Suno, and Udio audio.
- `https://aidetector.ceptile.com/ai-checker-for-teachers` — FERPA-compliant student paper review with ESL-safe calibration.

---

## 2. Text Detection Engine: The 46 Linguistic Signals

The text detection engine calculates 46 discrete linguistic and statistical signals organized across three core diagnostic pillars:

### Pillar I: Syntactic Regularity (Structural Predictability)
1. **Perplexity Variance (PPL-V):** Measures the fluctuation of token prediction entropy across consecutive sentences. Human prose exhibits high entropy variance; LLMs produce flat, uniform token probabilities.
2. **Burstiness Skew (BST-S):** The standard deviation of sentence lengths divided by the mean. High burstiness is characteristic of human writing; low burstiness indicates machine pacing.
3. **Punctuation Rhythm Regularity (PRR):** Evaluates the cadence of commas, semicolons, em-dashes, and terminal punctuation. AI models favor repetitive clause lengths (12–16 words between commas).
4. **Sentence Length Coefficient of Variation (CV-SL):** Ratio of sentence length standard deviation to sentence length mean ($CV = \sigma / \mu$). Humans routinely vary between 3-word punchy sentences and 45-word compound clauses ($CV > 0.65$). LLMs cluster tightly around $CV \approx 0.30 - 0.45$.
5. **Clause Depth Uniformity (CDU):** Syntactic parse tree depth variance across clauses.
6. **Passive Voice Equilibrium (PVE):** Distribution of passive versus active voice constructions. AI prose exhibits hyper-stabilized passive voice ratios (14–18%).
7. **Transition Density Index (TDI):** Frequency of formal discourse transitional markers (e.g., "Moreover", "Furthermore", "In conclusion", "It is important to note").
8. **Subordinate Clause Stacking (SCS):** Ratio of dependent clauses nested within complex sentences.
9. **Syllabic Rhythm Entropy (SRE):** Shannon entropy of syllable counts per word in 100-word sliding windows.
10. **Terminal Punctuation Entropy (TPE):** Variation in punctuation endings (period, question mark, exclamation, semicolon).
11. **Sentence Starter Redundancy (SSR):** Repetition of initial parts of speech (e.g., beginning 5 consecutive sentences with a gerund or subject pronoun).
12. **Parenthetical Frequency (PF):** Occurrence of parenthetical statements, em-dashes, and qualifying remarks. Humans use parentheticals irregularly; AI uses them systematically.
13. **Paragraph Flow Monotony (PFM):** Sentence count and word count uniformity across consecutive paragraphs.
14. **Grammatical Symmetry Factor (GSF):** Parallel structure adherence across paired coordinate clauses.
15. **Coordination Conjunction Ratio (CCR):** Proportion of sentences beginning with coordinating conjunctions ("And", "But", "So").

### Pillar II: Lexical Predictability (Vocabulary Distribution)
16. **Zipf Rank Divergence (ZRD):** Deviation of word frequency rankings from Zipf's empirical power law ($f(r) \propto 1/r^s$). Human vocabularies conform closely to $s \approx 1.0$; LLMs truncate the long tail of rare words ($s > 1.25$).
17. **Moving-Average Type-Token Ratio (MATTR-50):** Sliding 50-word window vocabulary diversity ($V / N$). Prevents text-length distortion present in standard TTR.
18. **Function Word Dispersion (FWD):** Uniformity of high-frequency closed-class grammatical words (determiners, prepositions, auxiliary verbs).
19. **Content Word Rarity Ratio (CWRR):** Proportion of vocabulary appearing in the bottom 5% of English usage corpora.
20. **Hedging Density (HD):** Frequency of epistemic qualifiers (e.g., "crucial", "testament", "delve", "beacon", "pivotal", "foster", "landscape", "underscore").
21. **Adjective-to-Noun Ratio (ANR):** Ratio of descriptive adjectives to nominal targets. AI writing displays an inflated, balanced modifier density.
22. **Verb Variety Index (VVI):** Diversity of action verbs versus repetitive copular verbs ("is", "are", "serves as").
23. **Collocation Predictability (CP):** Likelihood of two-word idiomatic combinations conforming to top-ranked trigrams.
24. **Bigram Surprise Metric (BSM):** Negative log probability of adjacent token transitions.
25. **Thesaurus Synonym Replacement Spike (TSRS):** Detection of artificially substituted synonyms characteristic of paraphrasing and humanizer tools (e.g., QuillBot, Undetectable AI).
26. **Repetition Penalty Score (RPS):** Penalty triggered when exact 4-gram or 5-gram phrases recur within a 200-word span.
27. **Hapax Legomena Ratio (HLR):** Frequency of words that appear exactly once in the entire document.
28. **ESL-Safe False-Positive Damper:** Specialized adjustment attenuating rare vocabulary penalties for Non-Native English (ESL) writers, who often use formal, standardized phrasing without machine origin.
29. **Latinate vs Anglo-Saxon Lexical Balance:** Proportion of Latinate multisyllabic terms versus Germanic roots.
30. **Nominalization Index (NI):** Frequency of abstract nouns derived from verbs (e.g., "-tion", "-ment", "-ance").

### Pillar III: Contextual Stability & Semantic Coherence
31. **Semantic Vector Drift (SVD):** Continuity of contextual topic embeddings across adjacent paragraph boundaries.
32. **Hallucination Anchor Scan (HAS):** Identification of generic, non-verifiable claims and pseudo-citations.
33. **Register Whiplash Metric (RWM):** Sudden shifts between hyper-formal academic prose and colloquial conversational phrasing (common in human-edited AI outputs).
34. **Topic Sentence Alignment (TSA):** Mathematical degree to which paragraph introductory sentences summarize following sentences.
35. **Information Density Per Token (IDPT):** Propositional content per 100 words. AI text frequently dilutes factual propositions with rhetorical filler.
36. **Contradiction Variance (CV):** Measurement of local thesis consistency across distant paragraphs.
37. **Listicle Structuring Bias (LSB):** Probability of hierarchical bullet lists, bolded introductory phrases, and symmetrical item lengths.
38. **Conclusion Tropes Detector (CTD):** Detection of formulaic closure constructions (e.g., "In conclusion", "As we look to the future", "Ultimately, it remains to be seen").
39. **Rhetorical Question Frequency (RQF):** Frequency of performative questions used as transitional devices.
40. **Quotation Integration Index (QII):** Contextual syntactical embedding of attributed quotations.
41. **Source Code Indentation Entropy (SCIE):** Indentation regularity, comment narration density, and identifier naming conventions for code analysis.
42. **Code Comment Redundancy (CCR):** Over-explanation of obvious code logic (e.g., `// loop through the array`).
43. **Code Variable Naming Entropy (CVNE):** Symmetrical camelCase or snake_case token length distributions.
44. **Context Gain Ratio (CGR):** Measurement of information gain as context length increases.
45. **Swap Fragility Metric (SFM):** Sensitivity of document semantics when adjacent sentence pairs are transposed.
46. **Global Ensemble Prior (GEP):** Calibrated logistic regression combining all 45 feature dimensions into a normalized probability score $P(\text{AI}) \in [0.0, 1.0]$.

---

## 3. Image Forensics Engine

The image detection engine parses JPEG, PNG, WebP, and AVIF files locally using Canvas and ArrayBuffer decoders:

1. **Error Level Analysis (ELA):** Resaves the image at a known 95% compression baseline and computes the pixel-level difference matrix. Synthetic images generated by diffusion models (DALL-E 3, Midjourney v6, Flux, Stable Diffusion XL) exhibit uniform error levels across high-contrast edges and textures, whereas authentic camera photographs display distinct error gradients between focal subjects and backgrounds.
2. **Bayer CFA Demosaicing Residuals:** Real digital camera sensors use a Color Filter Array (Bayer pattern RGGB) and interpolate color channels via demosaicing algorithms. Authentic raw/camera captures leave periodic green-channel interpolation fingerprints. Fully synthetic AI images generated by neural decoders lack optical CFA interpolation patterns.
3. **Discrete Cosine Transform (DCT) Noise Grids:** 8x8 block-level discrete cosine transform coefficients analysis. AI generative models generate subtle grid-like frequency artifacts due to convolutional upsampling and deconvolution layers.
4. **Metadata & C2PA Provenance Parser:** Extracts EXIF, XMP, IPTC, and Coalition for Content Provenance and Authenticity (C2PA) cryptographic assertions. Inspects for `Software`, `Prompt`, and `Model` markers (e.g., Midjourney, DALL-E, Adobe Firefly) while identifying stripped metadata typical of adversarial cleaning.
5. **Splice & Composite Grain Variance:** Computes 4x4 spatial quadrant variance ratios to identify composited or in-painted regions where local noise floors diverge from the primary frame.

---

## 4. Video Forensics Engine

The video detection engine samples frames across temporal boundaries using the HTML5 Video element and Canvas pixel extraction:

1. **Temporal Optical Flow Continuity:** Measures motion vector consistency across adjacent frames. Generative video diffusion models (OpenAI Sora, Google Veo, Kuaishou Kling, Runway Gen-3, Luma Dream Machine) frequently suffer from temporal coherence decay: background objects mutate, limb geometries flicker, and physical gravity vectors drift over spans of 24–48 frames.
2. **Inter-Frame Shot Noise Variance:** Authentic camera sensors exhibit continuous physical sensor noise across sequential frames. AI generated video frames often exhibit static noise patterns or sudden frame-to-frame noise drops characteristic of latent space interpolation.
3. **Diffusion Motion Blur Geometry:** Analyzes directionality of motion blur. Physical cameras produce linear motion blur along velocity vectors; diffusion video synthesizers produce isotropic gaussian blur lacking directional shear.

---

## 5. Audio & Voice Forensics Engine

The audio detection engine utilizes the Web Audio API (`AudioContext`, `AnalyserNode`) to evaluate raw PCM audio waveforms:

1. **Spectral Centroid & Bandwidth Tracking:** Fast Fourier Transform (FFT) analysis tracking spectral centroid movements. Synthetic voice models (ElevenLabs, Bark, Tortoise) display unnatural spectral shelf plateaus in the 3 kHz to 7 kHz range.
2. **Neural Vocoder Phase Coupling:** HiFi-GAN, WaveGlow, and BigVGAN vocoders synthesize audio waveforms from mel-spectrograms. This phase reconstruction process introduces subtle harmonic phase lock anomalies absent in natural human vocal tract acoustics.
3. **Synthetic Silence Profiling:** Evaluates ambient floor noise during pauses. Human voice recordings contain room tone and respiratory microphone acoustics; AI audio contains digital mathematical zero-amplitude silences or artificially injected stationary pink noise loops (Suno, Udio).

---

## 6. Empirical Benchmark Evaluation

The platform's scoring calibration is derived from rigorous cross-validated testing against a balanced empirical corpus of 2,743 real-world documents spanning news journalism, student academic essays, creative fiction, technical reports, and synthetic outputs from GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3, Mistral Large, and DeepSeek:

| Metric | Score | Explanation |
| :--- | :--- | :--- |
| **ROC AUC** | **0.966** | Area under the Receiver Operating Characteristic curve across mixed domains. |
| **Overall Accuracy (at 60% Threshold)** | **90.5%** | Correct classification rate on held-out test splits. |
| **False-Positive Rate (FPR at 60%)** | **6.2%** | Rate of human text incorrectly flagged as AI. |
| **False-Positive Rate (FPR at 70%)** | **4.6%** | Conservative threshold recommended for academic integrity evaluations. |
| **Sensitivity / Recall (AI Content)** | **92.1%** | Proportion of true AI documents successfully identified. |
| **Non-Native English (ESL) Precision** | **94.8%** | Maintained via our specialized ESL-Safe False-Positive Damper. |

---

## 7. Guidelines for Educators & Academic Institutions

AI Detector by Ceptile provides specialized guidance for teachers, professors, and academic integrity officers:

1. **Never Expel or Penalize Solely on a Score:** A detector score is statistical evidence of machine-like syntactic regularity, not forensic proof of academic misconduct. False positives can occur when human authors write in highly structured, formulaic, or standardized formats.
2. **Mandatory Human-in-the-Loop Review:** Scores must always be accompanied by a human review of the student's research notes, writing history, rough drafts, and oral defense of their work.
3. **Formative Rather than Punitive Application:** Use the tool as a pedagogical teaching aid to identify over-reliance on formulaic prose, encourage authentic voice, and guide students on responsible citation practices.
4. **FERPA & Student Privacy Compliance:** Because the tool runs 100% locally in the browser with zero cloud storage, uploading or pasting student papers does not violate FERPA (Family Educational Rights and Privacy Act) or student copyright.

---

## 8. Legal, Privacy & Corporate Transparency

- **Legal Entity & Ecosystem:** Developed and operated by Ceptile (https://ceptile.com), founded by Shaurya Agarwal.
- **Privacy Policy:** Strictly zero-upload. Zero tracking cookies, zero analytics beacons, zero server-side storage of user inputs. Conforms to GDPR, CCPA, VCDPA, and COPPA standards.
- **Terms of Service:** Users retain 100% intellectual property ownership of their analyzed content. Ceptile claims 0% rights or licenses over submitted text.
- **Contact Channels:**
  - General & Support: `hello@ceptile.com`
  - Security Vulnerability Reporting: `security@ceptile.com`

---
*End of Complete LLM Knowledge Base Dump.*