How we research
- Controlled samples. Each experiment starts from known files with verified metadata, not random downloads.
- Published method. Every study lists the exact steps, tools, versions and settings so results can be replicated.
- Raw observation. We publish what was measured — including values that don't fit the story.
- Calibrated conclusions. Experimental observations apply to the tested versions and paths. We do not generalize them into universal claims about platforms.
Studies
| Study | What it tests | Status |
|---|---|---|
| What Happens to EXIF When a Photo Goes Through WhatsApp? | Controlled before/after test: dimensions, size, EXIF, GPS, structure, hashes. | Planned |
| Which Metadata Survives an Instagram Upload? | Feed, story and DM paths compared against the original file. | Planned |
| Original Photo vs Screenshot of the Same Photo | What the camera records that the screenshot pipeline erases. | Planned |
| JPEG Recompression: What Each Re-Save Changes | Ten sequential re-encodes tracked through quality, structure and hashes. | Planned |
| Reading C2PA Test Files with Known Provenance | Marker detection behavior on documented sample credentials. | Planned |
| Camera Original vs Edited Copy: Signal Divergence | Controlled edit pass and what it leaves behind. | Planned |
Study pages are published as each experiment completes — with its method, data tables and screenshots. We won't publish summaries without the underlying observations.
Example protocol: the WhatsApp study
To show what "published method" means, here is the full protocol for the first study. Results publish only when the run is complete.
| Step | Action | Recorded |
|---|---|---|
| 1 | Start with 5 camera-original JPEGs of known content, hashes and full EXIF/GPS (recorded with consent, GPS set to a test location) | SHA-256, full metadata, structure map |
| 2 | Send each via WhatsApp normal photo send (Android and iOS clients, current app versions documented) | Client/OS/version |
| 3 | Save the received image (chat media download, not screenshot) | SHA-256, full metadata, structure map |
| 4 | Repeat each send in Document mode as the control arm | As above |
| 5 | Analyze original vs received pairs on PictureMatters; diff every field | Dimensions, size, EXIF delta, GPS, timestamps, quality estimate, entropy profile, dHash distance |
Observations will apply to the tested app versions and dates — not as universal claims about WhatsApp. Official platform documentation will be cited alongside, and any divergence between the two will be called out explicitly.
Authoritative references we rely on
- C2PA Technical Specification — provenance manifests and Content Credentials.
- IPTC Photo Metadata Standard — including digital-source-type for AI labelling.
- ExifTool documentation — the reference implementation for EXIF/XMP field behavior.
- W3C JPEG overview and the ITU-T T.81 specification — JPEG marker and quantization behavior.
- ISO 12232 / EXIF standards — camera metadata field definitions (CIPA DC-008).
- Platform behavior (WhatsApp, Instagram): cited from official documentation where available; otherwise labeled explicitly as our measurement with versions and dates.