What sits behind every verdict — the complete pipeline, no black box.

Four bands, no black-and-white verdict.

Every verdict is bound to a confidence value (0–100). The bands are chosen so the middle (SUSPICIOUS) is a real call to inspect, not a comfortable evasion.

Five forensic layers.

Local checks first, external APIs only when the fast detectors cannot decide. This ordering returns most results in seconds and rules out single-signal bias.

1. 01

   Metadata Forensics

   EXIF, MakerNote, and IPTC traces an image leaves on capture or edit.

   EXIF ForensicsExifTool

   Reads camera metadata, software tags, GPS, and timestamps. Surfaces inconsistencies between claimed capture context and actual sensor traces.

   iPhone MakerNote AnalysisApple MakerNote

   Parses Apple-specific metadata fields (Live Photo, HEIC pipeline, Lens ID). A complete, valid MakerNote is a strong authenticity signal.

   IPTC Digital Source TypeIPTC NewsCodes

   Reads the official IPTC NewsCodes tag XMP-iptcExt:DigitalSourceType. Editing software writes here itself whether an image is purely captured, algorithmically augmented, or fully generated.

   AI Editor FingerprintsClaimScan

   Custom pattern matcher for editor-specific traces in the XMP Credit/CreatorTool fields: Apple Photos Clean Up, Adobe Firefly, Google Magic Eraser, Samsung Object Eraser.

2. 02

   Image-Level Forensics

   Pixel and compression traces that surface post-hoc editing.

   ELA Compression AnalysisError Level Analysis

   Compares compression levels across image regions. Inserted or edited areas often show diverging error levels.

   Sensor Noise Profilesharp

   Measures sensor noise across image regions. Native captures show a self-consistent noise pattern; composites rarely do.

3. 03

   AI Generation Detection

   Three independent detectors hunting for diffusion, GAN, and vision-LLM patterns.

   Specialised AI Detector

   Specialised classifier trained on diffusion and GAN generation fingerprints from a large reference corpus.

   Independent AI Forensics

   A second multi-modal classifier with an independent training base. Runs in parallel — different models surface different generators.

   Vision Consistency Model

   Reviews physical plausibility, shadow and material consistency at the image level, and contributes a reasoned second opinion.

4. 04

   Cross-Reference

   Finds reused and web-known images via hashing and reverse search.

   Perceptual Hash Matching64-bit DCT pHash

   Computes a 64-bit DCT hash per image and matches identical or lightly modified reuse.

   Reuse AnalysisClaimScan

   Matches the pHash against a broad reference set, privacy-safe and without sharing images. The same image surfacing elsewhere is a strong reuse signal.

   Reverse Image Search

   Web-wide reverse image search. A web hit points to stock photo theft or open-web reuse.

5. 05

   Behavioral & Contextual

   Case level: multiple images, multiple submissions, same person — patterns a single image cannot reveal.

   Capture Timeline Analysis

   Cross-checks capture times across the images of a case for plausibility.

   GPS Plausibility Check

   Checks GPS coordinates against claimed capture location, altitude, and range.

   Device Chain Analysis

   Surfaces device switches within a case (e.g. iPhone and Android in the same submission).

   Visual Similarity Clustering

   Clusters similar images across cases to surface reuse patterns.

   Hit & Run Pattern Detection

   Finds repeated claim patterns from the same person across submissions — including very short time deltas.

Source: config/forensicStack.json — the same file that powers the stack diagram on the homepage and the marketing-claims audit. Visualisation on the homepage.

How findings become a verdict.

Every detector returns a probability, not a truth. The aggregation logic combines them across three axes — severity, confidence, direction — and is step-by-step traceable in the PDF report.

The exact per-step formula appears in every PDF report. Want it without an example case? We send the spec by email on request.

What we cannot do.

Forensic tools that claim to catch every fake are lying. This list states openly where Claimscan hits its limits — so you can place the verdict in context.

Where your images live — and how long.

Hetzner Cloud Nuremberg, own Postgres, Redis and MinIO instances — no US hyperscalers. Auto-delete windows depend on plan: 30 days (Free), 90 (Starter), 180 (Growth), 365 (Enterprise). Perceptual hashes persist indefinitely — they enable privacy-safe reuse detection and cannot be inverted into an image.

GPS data is used only for the plausibility check against the shipping address — it never lands in a persistent table. DPA available as PDF in-app, signed by the sole proprietorship of Piyal Ranasinghe.