WELCOME TO YEHUDA

AI Model Current Capabilities

Updated: December 09, 2025 – Please check often www.yehuda.com for updates.

Please make sure that no tags or paper are visible, as they may reflect light and cause the AI to give an incorrect result.

💎 Loose Diamonds

  • Minimum size: Greater than 0.25 carat
    • Divided tray (preferred): Place one diamond per square.
    • Undivided tray: Ensure stones are separated and do not touch each other.
  • Best practice: Place diamonds table-down for optimal results.

💍 Mounted Jewelry 

  • Minimum mounted stone size: Greater than 0.40 carat
  • Supported jewelry types:
    • Simple designs only, such as solitaire rings or stud earrings.
    • Stones must not touch each other.
  • Bezel-set jewelry is supported but may have reduced detection accuracy.
  • Placement: Ensure stones face upward for accurate detection.

⚠️ Important Notes & Limitations 

  • Detection misses:
    • Occasional misses may occur, especially in jewelry.
    • If a stone is not detected (not painted by the AI model), slightly adjust the piece and run the test again.
  • False positives:
    • The AI may occasionally mark non-stone areas (metal parts or reflections). These markings should be disregarded.
    • These markings should be disregarded.
  • Detector vs. AI accuracy:
    • Yehuda detectors will not miss a lab-grown diamond when the final decision is verified by a human.
    • A small percentage (about 2%) of natural diamonds are detected as LGD (false positives) by the detector. In those cases, the AI model will also classify them as LGD, because it relies on the detector’s physical reading.
  • AI accuracy:
    • Currently about 99% in our internal tests, and the model is still learning and improving.
  • Note:
    • The AI model serves as an advisory tool, not a final authority.
    • Continuous updates and training will further enhance its performance.

📸 Sample Images 

Here is a “heat map” table showing the current internal accuracy of the AI model.

The AI model confidence threshold is currently set to 80%. For every stone, the AI provides a confidence score for its prediction. We can choose the minimum confidence level at which a stone will be classified; any stone below that threshold is not classified (painted in black) If we raise the confidence threshold, we will get fewer stones classified, but with higher certainty.

The rows represent the actual stone type (ground truth), and the columns represent the AI model’s prediction.

How to Read This Table:

This table shows both the accuracy and the exact types of mistakes the AI model makes in our internal tests on a sample of 14,112 stones, when it is set to mark stones with 80% confidence.

For example, look at the Moissanite row (How were the Moissanite stones predicted by the AI model)):

  • 1.02% of Moissanite stones were classified as NATURAL.
  • 0.00% of Moissanite stones were classified as CVD.
  • 0.00% of Moissanite stones were correctly classified as HPHT.
  • 1.11% of Moissanite stones were classified as CZ.
  • 97.87% of Moissanite stones were classified as MOISSANITE.

Using the same logic, you can read the other rows to understand both the overall AI accuracy and the pattern of errors.

Important Clarification
The percentages above describe only how well the AI model classifies stones in our internal testing environment, based on the data it receives. Real-life performance may differ slightly.