The examination, in full, before you pay for it.
This page is linked from every dossier we write. It is the setting: there is no depth dial, no model picker and no advanced options, because the method version is ours and it is published.
Layer 1 — Provenance
We parse the container ourselves: JPEG markers and quantisation tables, PNG chunks, RIFF chunks, ISO BMFF boxes. From those we read the encoder fingerprint, the estimated save quality, the EXIF and GPS fields, embedded XMP and text chunks, and any C2PA Content Credentials store. We then check the file's own claims against each other — does the camera it names agree with the software it names, does the capture time agree with the modification time, do the recorded dimensions agree with the pixels present. On files over about 1 MB we read the first 768 KB and last 384 KB, which is where container metadata lives.
Layer 2 — Synthesis signals
A frontier multimodal model examines the still, or frames sampled from the clip, against the fixed checklist below. It is instructed to name the object and the place in the frame, to return the bounding box its reading came from, and to answer "not testable" whenever the frame does not contain what a signal needs. It is forbidden to return a score. Each reading carries a strength: clear, or weak.
Layer 3 — Measurements
Arithmetic, run in your own browser on the decoded frame, never a model: the Laplacian residual across 48 tiles (grain), the mean gradient per tile (detail), blocking energy at each of the eight pixel phases (compression grid), and highlight and shadow clipping. The numbers are printed in the dossier so anyone can argue with them.
Layer 4 — Finding, next steps and limits
The finding is resolved by a fixed rule over the signals, not by averaging opinions. Signs of synthesis requires either a declared generator string in the file's own metadata, or two signals reading anomalous at clear strength. Consistent with camera capture requires no anomalous signal at all and at least six consistent readings. Everything else is Inconclusive, which is a finding and not a failure. We then write the reporting moves that would actually settle the file, what would change the finding, and what we could not test.
The checklist — 17 signals
run in this order- 01
Text, lettering and signage
synthesisfree triage- establishes
- Whether letterforms, kerning and repeated glyphs behave the way printed or rendered text does in a photograph.
- cannot
- Establish anything when the frame contains no text, and it is defeated by text that has been inpainted correctly.
- 02
Hands and extremities
synthesis- establishes
- Finger count, joint articulation, nail structure and the way a hand meets an object it holds.
- cannot
- Say anything about a frame with no visible hands, or one where hands are motion-blurred.
- 03
Teeth, eyes and ears
synthesis- establishes
- Tooth separation and count, pupil and catchlight geometry, iris detail, and ear cartilage structure.
- cannot
- Work below roughly 100 pixels of face width, or through heavy compression.
- 04
Hair and subject boundaries
synthesis- establishes
- Whether strand-level detail continues into the background the way hair photographed by a camera does, and whether the subject outline has been cut.
- cannot
- Distinguish a synthetic boundary from an aggressive portrait-mode cutout on a camera-captured photograph.
- 05
Lighting direction and shadows
synthesis- establishes
- Whether every shadow in the frame agrees on where the light is, and whether shadow softness matches the apparent source size.
- cannot
- Resolve scenes with several light sources, overcast light, or no cast shadows at all.
- 06
Reflections and speculars
synthesis- establishes
- Whether reflective surfaces — eyes, glass, water, metal, screens — return the scene that is actually in front of them.
- cannot
- Do anything in a frame with no reflective surface.
- 07
Depth of field and focus falloff
synthesis- establishes
- Whether the plane of focus and the way blur increases with distance are consistent with one lens and one aperture.
- cannot
- Separate a synthetic depth map from a phone's computational portrait mode.
- 08
Skin and fabric micro-texture
synthesis- establishes
- Whether pore, wrinkle and weave detail is present at the resolution the frame claims, and whether it varies naturally across the subject.
- cannot
- Survive heavy beauty filtering, denoising or a low-resolution re-encode, all of which erase the same detail.
- 09
Edge quality and halos
synthesis- establishes
- Whether object edges carry the sharpening halos, chromatic fringing and per-edge softness a camera's optical path leaves.
- cannot
- Be read on an image that has been re-sharpened after download.
- 10
Repeated structure and tiling
synthesis- establishes
- Whether patterns in the frame — crowd faces, foliage, bricks, fabric print — repeat more exactly than a physical scene repeats.
- cannot
- Flag manufactured objects that genuinely are identical.
- 11
Physical plausibility and scale
synthesis- establishes
- Whether object sizes, contact points, gravity, water, smoke and crowd behaviour hold together as one physical scene.
- cannot
- Rule out an unusual but genuinely photographed event, which is exactly the kind of image people doubt.
- 12
Perspective and background continuity
synthesis- establishes
- Whether vanishing lines agree, and whether structures continue correctly behind the subject.
- cannot
- Read wide-angle or corrected lenses reliably, and it is defeated by shallow depth of field.
- 13
Temporal stability across frames
synthesis- establishes
- Whether identity, garment detail, background objects and lighting persist frame to frame, and whether flicker appears where motion does not explain it.
- cannot
- Run on a still image, and it is weakened when we sample only a few frames of a long clip.
- 14
Sensor noise uniformity
measurement- establishes
- Whether the high-frequency residual is spread across the frame the way one sensor at one ISO spreads it.
- cannot
- Distinguish a spliced region from a region that was simply denoised, and it is close to useless on a heavily compressed file.
- 15
JPEG block grid alignment
measurement- establishes
- Whether 8-pixel blocking artefacts sit on the file's own grid, which is what a single JPEG save leaves.
- cannot
- Count how many times a file has been saved, or tell a crop from a paste.
- 16
Resampling and interpolation
measurement- establishes
- Whether local detail is smoother than the file's stated resolution should allow, which is what upscaling leaves behind.
- cannot
- Tell an upscaled camera photograph from an upscaled synthetic one.
- 17
Tonal range and clipping
measurement- establishes
- Whether highlight and shadow clipping look like a sensor's response or like a rendered image's.
- cannot
- Carry weight on its own; it is context for the other measurements.
Known failure modes
- ·A generated image that carries no metadata and no visible defect will read as Inconclusive. We will not guess to be useful.
- ·A camera photograph that has been upscaled, denoised, beauty-filtered or re-encoded by a platform loses the same detail synthesis lacks, so its measurements weaken. We say so rather than counting it against the file.
- ·We do not yet validate C2PA signatures cryptographically. We report that a manifest is present and what it declares.
- ·We sample up to 6 frames of a clip, so a defect confined to frames we did not sample will be missed. The dossier states the sample.
- ·We do not examine audio at all, and we will not guess at it.
- ·Detection accuracy against any single generator decays as that generator improves. The method version on every dossier exists so a dossier can be read years later against what we actually ran.
Changes
- 2026-08-30
- 2026.08 — first published method. 17 signals across four layers, three findings, no percentage.