CardsBenchmark

Method

Exact printing or it's wrong. No partial credit. Every score comes from a run.

Identification and recognition

CardsBenchmark follows CREPS (Card Recognition Evaluation Protocols). Identification is top-1 on the card. Recognition is top-1 on the exact printing: set, card number and printing. We report both side by side and never merge them. The leaderboard ranks recognition.

Scanners method v0.1

Consumer scanner apps, scored S C A N. An app ranks once 4 dimensions are scored.

LetterDimensionTestScored by
SFirst answerOpen cold, scan card #1reviewer
S20 cardsScan 20 public-set cards back to backreviewer
AExact printingCREPS public set (computed)computed from runs
AParallel / variant trapFrozen card #2 + the siblings stratumreviewer
ABase-era vintageFrozen card #3 + the vintage stratumreviewer
AJapanese and other languagesThe language stratumreviewer
ABacks, fakes, customsThe far-OOD stratumreviewer
ASame photo twiceEvery public-set photo run twicereviewer
AGlare, sleeve, slab, blurThe hard stratumreviewer
CPrice with its sourceFrozen card #1 price screenreviewer
CCollection and exportAdd 4 cards, exportreviewer
CCondition estimateFrozen card #4reviewer
NPaywall before valueTap until money is asked forreviewer
NCaps and quotaBurn through the free tierreviewer
NClaims vs measuredVendor accuracy or speed claim vs our runreviewer

Graders method v0.1

AI grading apps, scored G R A D E under COGS limits. An app ranks once 4 dimensions are scored.

LetterDimensionTestScored by
GGroundingPillars measured and shown: centering, corners, edges, surface.reviewer
RRepeatabilityR1 within ±1: the same card, removed and re-shot three times.reviewer
AAgainst outcomesAgreement on raw photos of cards later graded professionally. Needs the truth set.reviewer
DDeliveryTime to a grade, and cost.reviewer
EEvidence and claimsThe No BS axis for graders: claims, paywalls, receipts.reviewer

Engines and APIs method v0.1

Recognition engines, scored under CREPS.

No dimensions defined in the database.