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.
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
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.
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.