BAO Methodology & Validation Center

BAO Grade is a photographic pre-grading system for trading cards. This center documents what the deployed system measures, what it cannot establish from photographs, how new evidence is promoted into grading behavior, and how professional returns are collected for validation. The goal is to make the system inspectable rather than asking collectors to trust a single unexplained number.

Current deployed versions

Live professional-return benchmark

Current validation status: collecting verified outcomes. BAO intentionally suppresses accuracy percentages while the verified sample is small. Showing a percentage from only a few returns would create false precision and would not support a responsible accuracy claim.

What BAO measures today

The production grading workflow evaluates four condition dimensions: centering, corners, edges and surface. The card plane is detected and rectified before measurements that depend on geometry, reducing the effect of camera perspective. Corner and edge evidence is evaluated across front and back photographs. Surface analysis can use optional directional-light photographs, especially for reflective or foil cards, but those extra views affect conclusions only when alignment and lighting-diversity checks support the comparison.

Evidence has to earn grade impact

BAO separates corroborated findings from inspection candidates. A new detector is not allowed to punish a card merely because it sees an anomaly once. Current perimeter logic requires agreement between independent photographic evidence channels before a new finding can lower a grade. Experimental reference-comparison and few-shot anomaly channels remain shadow-only until their own benchmark and promotion requirements are met. This conservative design is intentional: uncertain evidence should create a recapture or inspection prompt instead of a false defect.

Capture quality is part of the result

A photograph can hide or imitate damage. Glare can look like whitening, a slab can introduce scratches that are not on the card, and perspective can make an evenly printed card look off-center. BAO therefore treats image quality, boundary confidence and recapture instructions as part of grading rather than as cosmetic warnings. The optional Forensic Capture workflow adds directional front and back views and a reflective tilt view for cards that deserve closer surface inspection. Standard two-photo grading remains available.

What BAO does not claim

BAO does not authenticate a card, encapsulate it, guarantee a resale value, or claim that a BAO number is equivalent to a PSA, CGC, BGS, TAG or SGC grade. Different grading companies have different standards and judgment processes, and a photograph cannot reveal every physical characteristic available during direct inspection. The submission planner on a confirmed card uses raw market information and the collector’s own current cost and sale assumptions; it does not apply a hidden grade multiplier or predict a third-party grade.

How professional returns enter the validation program

After a card comes back from a supported professional grading service, its owner can submit the returned grade and certification number from the BAO report. That outcome is stored as collector-reported and unverified. It cannot automatically train the model and does not count toward public accuracy statistics. An operator must independently verify the professional outcome before it joins the validation set. Verified outcomes preserve the BAO grade and engine version that produced the original estimate so later model releases do not rewrite history.

How model releases are evaluated

Free-service scans can enter a private model-development review queue under the required photograph-use license. Predictions are not ground truth. Training examples require rights checks, annotation and validation. Captures of the same physical card should stay in the same dataset split so a model is not tested on near-duplicates of images it already saw during training. Candidate changes should be evaluated against frozen holdouts and promoted only when they improve useful sensitivity without creating unacceptable false-positive damage calls.

Why BAO waits before publishing accuracy percentages

Small samples can produce impressive-looking but unstable percentages. BAO therefore requires at least 30 independently verified professional returns before its public endpoint exposes numerical agreement rates. Even after that threshold, the metrics are descriptive: exact agreement, agreement within half a point, agreement within one point and mean absolute numerical difference. Results should also be examined by professional grading company and capture mode because combining different standards can hide meaningful differences.

Versioning and accountability

The trust layer records the deployed grading engine, capture mode, evidence schema and validation state. Product funnel events are first-party and privacy-minimized: analytics session identifiers are one-way hashed server-side and the dedicated product-event table does not store raw IP addresses. BAO can therefore measure where collectors abandon or complete the workflow without turning product analytics into another source of grading labels.

How BAO grades cards · Condition checker · Centering calculator · PSA pre-grading checklist

Grade your card free →