The model card

The full pipeline, measured on the test set

321 items never seen in training: 164 counterfeits and 157 authentic. Out-of-scope photos and the uncertainty zone count as “inconclusive”.

Counterfeits flagged “high risk”

[73.3% – 85.4%]79.3%

Counterfeits rated “low risk”

[1.3% – 7.5%]4.3%

Authentic items flagged “high risk”

[0.6% – 6.2%]3.2%

Authentic items rated “low risk”

[74.2% – 86.5%]80.3%

Correctness of “high risk” verdicts

[92.8% – 99.2%]96.3%

Correctness of “low risk” verdicts

[90.7% – 98.4%]94.7%

Inconclusive analyses

[12.5% – 20.6%]16.5%

Accuracy when the model decides

[92.9% – 97.8%]95.5%

[ … ] 95% CI

Without abstention

If the model were forced to decide every time, with no uncertainty zone:

Area under the ROC curve

[94.1% – 98.2%]96.3%

Forced accuracy (50% threshold)

[88.2% – 94.4%]91.3%

Best model published by the authors

95.0%

Human experts, according to the dataset authors

89.8%

These two references are not directly comparable to our measurements: the authors evaluate image by image, yet the dataset contains 116 exact duplicate files. Our split keeps every duplicate within a single subset, which makes the test stricter.

Calibration and thresholds

The displayed probability is recalibrated on the validation set (temperature scaling). The model returns “high risk” from 83.5% and “low risk” up to 20.0%; in between, it abstains.

Calibration error before

3.1%

Calibration error after

3.2%

Is the model looking at the item or the setting?

Probe: each photo is shrunk to a 24-pixel thumbnail, where no metal texture is readable any more, and a simple classifier is trained on it.

Accuracy from thumbnails alone

75.4%

Chance level

51.1%

Accuracy of a simple classifier on full photos

80.4%

Thumbnails alone do clearly better than chance: part of what the model learned comes from the context of the photos (background, framing, source) rather than the material. Its performance on your own photos may be lower than the figures above.

Scope gate

Before scoring a photo, the model checks that it resembles the jewellery it knows.

Test-set photos wrongly rejected

1.6%

Unrelated photos rejected (37 wallpapers)

37 / 37

Data

GoldNet dataset: smartphone photos, one photo per item.

Images

2,127

Licence

CC BY 4.0

Training

1,486

Validation

320

Test

321

Near-duplicate groups kept within a single subset

119

Reference — Z. Raisi, « GoldFormer: A Texture-Aware Vision Transformer-Based Algorithm for Detecting Near-Identical Images », Algorithms, 2026, 19(7), 530. DOI 10.3390/a19070530 https://github.com/zobeirraisi/GoldNet

Known limits

1. 01The result is a statistical estimate from a photo. It replaces neither a metal assay nor a professional's opinion.
2. 02The model learned from a single public dataset: GoldNet, 2,127 items.
3. 03The source does not document how the authentic / counterfeit labels were established.
4. 04The items come mostly from Iran and the Gulf market: reliability on other styles has not been measured.
5. 05Each item has a single photo in the dataset: combining several photos has not been validated experimentally.
6. 06The filter that rejects off-topic photos was only checked on a small informal probe.
7. 07The setting of the photos is partly enough to separate the two classes: the model may have learned context cues rather than material ones.