A forger copies the shape. Never the habits.
Our machine-learning comparison measures the habits: how each letter is built, how tall it stands next to the others, the slant, the spacing, where the pen joins and lifts. It sets the signature against genuine signatures by the same hand from the same years, so our examiners see the evidence, not just a likeness.
Stroke orderHow it was written, not just how it looks
From a single photograph, our model reconstructs the order and direction of every stroke. An examiner checks the reconstruction and corrects it where needed, then the model compares it with her genuine signatures.
A copy can match the shape exactly and still be written in the wrong order.
- 1Reconstructed by our model
- 2Checked by an examiner
- 3Compared with her genuine signatures
Drawing…
Seven questions, every time
The signature is isolated from the page and the model works through the same seven questions for every item.
Stroke order
Was it written in her order and direction, stroke by stroke?
Exemplar retrieval
Find the most relevant genuine signatures, contemporaneous ones first.
Structure
Letter construction, proportions, spacing, slant and connecting strokes.
Natural variation
Does it fall within the writer’s documented range?
Chronology
Does it match the style of the probable signing period?
Copies
Is it suspiciously close to one known signature, as a tracing would be?
Anomalies
Anything unusual is flagged for the examiner.
Exemplar retrievalCompare like with like
A hurried felt-tip signature at a signing table should be compared with hurried felt-tip signatures from the same years, not with a careful letter written twenty years later. The model finds the most relevant genuine signatures first, and contemporaneous ones come first of all.
- 1998 · Felt-tip
Signing table - 1996 · Felt-tip
Signing table - 1999 · Felt-tip
Dedication - 2001 · Felt-tip
Signing table - 1997 · Ballpoint
Signing table - 1998 · Fountain pen
Dedication - 2009 · Felt-tip
Signing table - 2019 · Felt-tip
Dedication - 2014 · Ballpoint
Signing table - 2004 · Ballpoint
Letter - 1984 · Fountain pen
Letter - 1989 · Ballpoint
Dedication
Same years, same pen, same situation: the closest thing to watching her sign that day.
Structural comparisonMeasure what the eye skips
A practised copy can look right at a glance. Measured, it drifts: capitals too tall, the slant too upright, the words too far apart, the pen lifted where the writer never lifts it.
The photograph: ink, paper, print showing through. Then the model separates the signature from everything else on the page.
Inside the range, or out of it
No one signs the same way twice, so the model learns each writer’s natural range, period by period. Drag the questioned signature around the map and see how the model reads each place.
It varies from her genuine signatures of the period the way they vary from each other. Consistent with the 1998 date.
- Nearest genuine signatures
- 1996 · Felt-tip
- 2002 · Felt-tip
- 2000 · Ballpoint
An illustration. The model measures many features at once; this map flattens them into two so the idea is visible. The writer and her signatures are made up.
What the examiner receives
The model produces an annotated comparison: the questioned signature beside the closest genuine examples, what supports it, what doesn’t, and anything unusual enough to need an expert’s eye.
Supporting
- Letter construction matches her 1995–2002 signatures
- Proportions and slant within her range for the period
- Fluent line, no hesitation at the joins
Discrepancies
- Flourish shorter than in most examples of the period
Flagged for the examiner
- The final “e” sits lower than in any example on file
An example with a made-up writer. Real comparisons use the writer’s own genuine signatures.
A tool for the examiner, not a replacement
The model
- Finds the most relevant genuine signatures on file
- Measures every one of them the same way
- Places the signature within, or outside, the writer’s range
- Flags copies that are too close and features that don’t fit
The examiner decides
- Whether similarities are meaningful or only superficial
- Whether differences are the writer’s natural variation
- Whether the line was written freely or imitated
- Whether inscription and signature were one act
- Whether anything contradicts authenticity
At level 1 a certificate rests on the model’s analysis of photographs, without a physical examination, and a person checks the result. From level 2 the item itself is examined in our lab, and at levels 4 and 5 one and then two independent experts form the opinion.
The model is only as good as its examples.
Every comparison rests on signatures we can trust. Building that library is the slowest, most careful part of our work.
The exemplar database