Skip to main content

AI analysis

No black box: here is exactly what the model does, what it relies on, and where its limits are.

In short

GradeMate sends photos of the front and back to a vision model, together with the published scale of the grading company you selected. The model assesses each face separately on four criteria — centering, corners, edges, surface — then produces a combined grade, a per-face list of detected defects and a confidence level. The analysis only covers what is visible in the image: it does not detect chemical alteration, recuts, or defects that require raking light or magnification.

How an analysis runs

  1. 1. The images are sent

    Front and back are attached to the request along with the target company's scale.

  2. 2. Each face is assessed

    The model processes front and back separately, so a good front cannot mask a damaged back.

  3. 3. Grades are combined

    The four criteria are aggregated into an overall grade, with identified defects and their estimated impact.

  4. 4. The result is validated

    The response is checked against a strict schema. If it is incomplete or inconsistent, the analysis fails explicitly instead of showing an invented result.

What the analysis detects well

Centering

Margin ratios are directly measurable in the image: the most reliable criterion for automated analysis.

Visible whitening

Soft corners, whitened edges and border chipping stand out clearly on a properly framed photo.

Obvious defects

Sharp creases, deep scratches, stains, print defects and layer misregistration.

What it cannot do

No photo-based tool can. Better to say so plainly.

See under raking light

Some bends and micro-scratches only appear when tilting the card under a point source. A flat photo misses them.

Detect alteration

Recuts, chemical cleaning, colour touch-ups: these require physical examination and specialised equipment.

Issue an official grade

Only an accredited company issues a grade and a sealed slab, and that is what the market values.

Frequently asked questions