Centering

Preamble

Centering takes front and back photos and turns them into measured borders, visible overlays, confidence, and grading-company impact. The useful answer is the one that shows exactly where the card was found and where the image stopped being trustworthy.

---

The Phone Has Already Bent The Card

The first slice is a local prototype with two photos for one card: front and back. By the time the app receives them, the phone has already chosen the light, angle, crop, compression, and distortion that the software must measure through.

Centering starts with the part of the card a photo can most plausibly show: the border. The app has to find the card, orient it, flatten the perspective, detect the outer and inner boundaries, calculate ratios, and translate those ratios into PSA, BGS, CGC, ACE, and TAG impact.

The Overlay Is The Claim

A ratio by itself asks for trust. An overlay lets the collector see the claim: the plane the app chose, the edges it detected, the border it measured, and the parts of the image that weakened the answer.

The geometry has to repeat. The same image should produce the same card plane, the same detected boundaries, the same border ratios, and the same confidence state. LLM-written code can help build the engine, but the measurement itself has to be pinned down by deterministic geometry, typed thresholds, test images, and visible failures.

Some Photos Should Fail

A compromised image should not produce a confident answer. Glare, blur, missing corners, sleeve obstruction, low resolution, bad crop, and warped perspective are measurement facts. They belong in the output as uncertainty, refusal, or a request for a better photo.

Video, live capture, mobile flow, and batch mode come later. Batch mode is blocked until the app can lock front and back photos to the same physical card, because a fast workflow that mismatches sides would be worse than a slow one.