BAO Grade is building its grading claims from versioned evidence rather than marketing estimates. This page is the public registry for studies we intend to publish as the verified sample grows. Planned studies stay marked as collecting evidence until their data and evaluation gates are sufficient.
Current validation status: Collecting evidence: 0 of 30 independently verified professional returns required before BAO publishes agreement percentages.
Accuracy Lab status: Blinded Accuracy Lab: 0 of 30 verified physical-card cases; repeatability: 0 physical cards across 0 captures.
Compare frozen BAO photographic estimates with independently verified PSA, CGC, BGS, TAG and SGC returns. Professional labels are revealed only after BAO captures are frozen, and withheld grades remain abstentions.
Repeated photographs of one physical card stay grouped as one sampling unit. BAO measures grade spread, standard deviation, abstentions, and pairwise agreement within 0.5 and 1.0 grade points.
Measure disagreement between raw-photo border geometry and the rectified physical card plane so camera angle is not mistaken for printed centering.
Measure whether directional-light corroboration reduces false surface findings on reflective cards while preserving sensitivity to real visible defects.
Measure detector-box precision separately from coarse region fallbacks. Approximate inspection regions are never scored or displayed as pixel-precise localization.
Track the confidence gap between top catalog candidates and require collector confirmation when the exact printing or variant remains ambiguous.
Professional outcomes must be independently verified before benchmark use. In the blinded Accuracy Lab, BAO captures are frozen before the professional label is revealed. Repeated photographs of the same physical card stay in one case and one dataset split so near-duplicates cannot inflate sample size or leak into a holdout. Benchmark cases are evaluation-only and are not automatically training labels. Capture failures and abstentions remain part of the measured system rather than being removed to improve reported accuracy.
When samples are large enough, BAO will publish exact agreement, agreement within half and one grade point, numerical bias, mean absolute difference, abstention rate, same-card repeatability across captures, and defect localization reliability. Results will be sliced by grading company, capture mode and other material conditions instead of combining incompatible populations into one flattering number.
A useful photographic pre-grader should be repeatable under ordinary phone capture, explain where its condition evidence came from, admit when the photograph cannot support a precise conclusion, and show how its estimates compare with independent professional returns. BAO is instrumenting those questions directly.
Methodology & validation · Card photography guide · How BAO grades cards