Confidence Score Algorithm
The confidence score expresses, from 0 to 100, how likely it is that the questioned signature comes from the same person as the references. The score is not the output of a single neural network; it is a weighted combination of the calibrated votes of 23 models trained with different architectures and on different datasets. That way no single model's blind spot can decide the result on its own.
How the Votes Combine
The models fall into three families: image-based siamese networks (8 models), sequence models working on the stroke skeleton (9 models) and classical classifiers using hand-defined biometric features (6 models). Each model compares the questioned signature with every reference separately and produces a calibrated probability between 0 and 1.
The consensus layer first takes the median of the per-reference results; this stops one badly scanned reference from dragging the score. It then weights the models by how well they did on an independent validation set. If one family clearly contradicts the other two, the score is automatically pulled towards the inconclusive band and the response is flagged with disagreement: true.
Score Bands
| Band | Score | Meaning | Recommended action |
|---|---|---|---|
| strong_match | 85–100 | Strong match; the large majority of models agree. | Continue the normal workflow. |
| inconclusive | 60–84 | Inconclusive; the evidence could point either way. | Ask for more references or refer to an expert. |
| mismatch | 0–59 | Mismatch; the signature is not consistent with the references. | Stop the transaction and examine the document. |
Band limits can be changed per team (Settings → Signature API), but the report always shows the default bands as well.
Explainability
Every response returns the features that contributed most to the score in the top_features array. The contribution value shows how many points that feature pushed the score up (positive) or down (negative); all contributions add up to the difference between the neutral starting point of 50 and the final score. For example, slant_angle: -9.6 says the slant angle deviates from the references and lowered the score by about ten points. The full list is in the feature matrix of the PDF report; for the feature definitions see Biometric Analysis.
What Lowers Confidence
With a single reference the person's natural variation cannot be measured; the score effectively tops out around 80. Send three to five references.
Scans below 300 DPI and heavy JPEG compression wipe out fine stroke detail; tremor and line-width measurements in particular suffer.
A wet-ink signature versus one made on a tablet, or a ballpoint versus a felt-tip, changes the line-width distribution. Use references of the same kind where you can.
Signatures change over the years. References dated close to the questioned signature give more reliable results.
The difference between 84 and 86 is not statistically meaningful. Always send scores close to a band limit to human review.