The "3R" Workbench: Reveal, Recover, Refine

Conventional AI detectors output a single unexplainable percentage and frequently guess on stripped files. OriginSift operates on a transparent Reveal → Recover → Refine workflow designed for journalists, researchers, educators, and everyday readers:

01

Reveal

We inspect cryptographic C2PA manifests, camera EXIF/XMP tags, DOCX/PDF revision history, invisible watermarks, multi-model neural classifiers, 2D FFT spectra, and 3×3 regional noise consistency—plus a 1-click OSINT Reverse-Search Launchpad when metadata is stripped.

02

Recover

When AI generators (AUTOMATIC1111, ComfyUI, SwarmUI, Fooocus, InvokeAI, NovelAI) embed workflow parameters in PNG chunks or JPEG/WebP EXIF headers, OriginSift recovers the exact prompt, negative prompt, checkpoint, LoRA modules, seed, and denoise mode.

03

Refine

For text and uploaded documents, OriginSift pairs multi-window neural classification and DistilGPT-2 perplexity with an honest Writing Style Coach that flags stock phrases and flat cadence—helping authors write more directly without claiming to "bypass" detectors.

How the Workbench Inspects Media & Documents

Every file submitted to OriginSift passes through deterministic provenance extraction first, followed by domain-specific signal analysis and calibrated multi-model neural triage:

🔐 Cryptographic C2PA, EXIF/XMP & Watermarks

Validates signed C2PA Content Credentials manifests against issuer certificates, parses camera EXIF and IPTC DigitalSourceType tags, and checks for public invisible watermark carriers. When social networks strip metadata, OriginSift computes SHA-256 and 64-bit perceptual dHash fingerprints for 1-click reverse image search across Google Lens, TinEye, Bing Visual Search, and Yandex.

🧩 Multi-Model Vision & Spatial Signal Triage

Combines a 3-model AI vision classifier ensemble (haywoodsloan/ai-image-detector-deploy, Organika/sdxl-detector, and umm-maybe/AI-image-detector) with Error Level Analysis (ELA), 2D Fast Fourier Transform (FFT) periodic grid detection, and a 3×3 regional Laplacian noise consistency grid to highlight localized splicing, face swaps, or diffusion upscaler artifacts.

🎙️ Audio & Video Temporal Inspection

Extracts evenly spaced OpenCV video keyframes with center-crop ROI checks, per-frame perceptual dHashes, and 1-click reverse frame search, alongside Short-Time Fourier Transform (STFT) Inferno spectrograms, fundamental pitch (F0) jitter tracking, and 4-second sliding-window Wav2Vec2 synthetic voice classification.

📄 Document Revision Provenance & Style Metrics

Reads OpenXML revision counts (cp:revision) and logged editing minutes (TotalTime) from Microsoft Word .docx files and incremental save markers (%%EOF) from .pdf files, alongside stock phrase density, sentence length variation (CV), Moving-Average Type-Token Ratio (MATTR) lexical diversity, and DistilGPT-2 token perplexity.

Evidentiary Standards & Why We Return "Inconclusive"

A core principle of OriginSift is separating verifiable cryptographic or container provenance from probabilistic statistical signals. Understanding the boundary between the two prevents both false accusations and misplaced confidence:

⚖️ Verifiable Provenance vs. Statistical Triage

When a file carries an intact C2PA cryptographic signature or embedded Stable Diffusion / ComfyUI generation parameters, its origin is deterministically verifiable. By contrast, pixel frequency analysis, audio spectrograms, and text perplexity are statistical indicators that can be influenced by heavy WhatsApp/X re-compression, studio lighting, or formal academic editing.

🧭 Three-Band Verdict Calibration

Instead of forcing a binary guess on degraded files, OriginSift classifies results into three transparent bands: Likely Authentic / Original, Inconclusive (Mixed or Stripped Signals), and Likely AI-Generated. Every indicator card displays its raw measurement, weight, and plain-English explanation.

Explore OriginSift Standalone Detectors & Field Guides

Open our dedicated single-purpose detectors or read our illustrated field guides:

🖼️ AI Image & Photo Detector

Check JPEG, PNG, and WebP images with our 3-model vision ensemble, 2D FFT spectral grid check, ELA, 3×3 splicing heatmap, and camera EXIF validator.

🎬 AI Video & Voice Deepfake Detector

Verify MP4/MOV/WebM videos and WAV/MP3 voice notes with a Frame-by-Frame AI Detection Timeline, temporal flicker forensics, and Wav2Vec2 audio windows.

✍️ AI Writing Style Checker & Sentence Coach

Inspect essays, articles, Word (.docx), and PDF documents for token perplexity, sentence burstiness, OpenXML revision history, and 1-click rewrites.

🧬 AI Prompt & ComfyUI Workflow Extractor

Recover embedded positive/negative prompts, seeds, samplers, checkpoint hashes, and ComfyUI node-graph JSON workflows from AI images.

📷 The Digital Birth Certificate: What is EXIF Data?

Plain-English guide using the Airport Passport Metaphor and custom diagrams to explain EXIF camera hardware tags, hollow AI headers, and social media stripping.

📸 5 Common Giveaways That a Viral Photo is AI-Generated

Step-by-step visual guide covering iris catchlights, hand geometry, background typography, C2PA/EXIF metadata checks, and 1-click OSINT reverse image search.

Engineering & Innovation

💡

Developed by Atanico Digital Solutions

OriginSift is engineered, designed, and maintained by Atanico Digital Solutions. We build transparent software architectures, provenance tooling, explainable machine learning systems, and enterprise digital solutions.

Visit Atanico Digital Solutions (atanico.com) ↗

Your Files Are Inspected in Memory and Not Stored

OriginSift processes uploaded images, audio, video clips, and documents in volatile memory to return your inspection response and does not store them on disk or in a database. User content is never used to train or fine-tune AI models. Review our Terms of Service and Field Guides for more details.