PristineGrader
Verified Tuesday, May 26, 2026

Same scan.
Same grade.
Every time.

PristineGrader is built on deterministic computer vision. Identical input always produces identical output — to the pixel, to the metric. We test it nightly. If it ever drifts, this page stops claiming it's verified.

Latest test run
Cards tested
5
Mix of Pokémon basic, full-art, and back designs
Runs per card
5
Each card processed end-to-end this many times
Divergences
0
Every field of every output identical across all runs
Method

bit-exact comparison of full pipeline output (centering, corners, edges, surface, quality)

The test rig lives in services/ml/scripts/test_repeatability.py. It runs the full pipeline N times across K test cards and compares the JSON output for every measurable field. If anything diverges, the script exits non-zero and this verification date doesn't update.

Why this matters

Other AI grading apps
5 different grades from 6 scans of the same card

App Store user review — same card, same lighting, same tripod

Most AI graders use models that have stochastic inference — small numerical differences across runs compound into different scores. You can't trust a number that changes when nothing changed.

PristineGrader
Identical output, every run, forever

Verified Tuesday, May 26, 2026 — 5 cards × 5 runs

Our pipeline is classical computer vision: no randomness, no model inference at the scoring layer. The output is a deterministic function of the input image. We test it because we owe it to you.

See it for yourself.

Upload the same card twice. Compare the reports. They'll be identical down to the last pixel measurement.