How to Grade Card Corners

Corners are the most damage-prone region of a trading card and the first place handling becomes visible; in most graded-card price structures, the corner subgrade moves value more than any other factor at the top of the scale. Yet corner evaluation has historically been the least consistent part of grading. Human graders weigh identical wear differently across days and individuals. Automated estimators introduce a second inconsistency: a model inferring condition from an uncontrolled photograph evaluates the photograph — its lighting, angle, and focus — as much as the card.

The methodology described here addresses both failure modes with a single principle: measure first, adjudicate second, and hold every judgment accountable to the measurement. Sections 3–6 describe the method, its published scale, the data assets that train and audit it, and how any issued grade can be independently verified. Section 7 states the method's limitations.

Table 1 — Terminology used throughout this document.

The Vision™ corner methodology is a five-stage pipeline. Figure 1 summarizes the flow; sections 3.1–3.5 specify each stage.

Figure 1 — The five-stage corner grading pipeline. Measurement (stage 3) precedes and constrains all model-based judgment (stage 4).

3.1 Calibrated capture

Both sides of the card are scanned on calibrated equipment under fixed lighting, distance, and resolution. Capture control is a methodological requirement, not an operational preference: variance in the input propagates to every downstream judgment, and no analysis stage can recover information the capture failed to record. Grading from uncontrolled photographs is out of scope for this methodology (see §7).

3.2 Corner isolation

Eight corners — four per side — are cropped into standardized examination windows. Each window preserves a traceable reference to the source scan, maintaining an unbroken chain from any corner-level finding back to the original capture.

3.3 Physical measurement

A proprietary silhouette measurement engine traces each corner's physical outline and computes geometric properties of the tip and converging edges: whether the outline is square, rounded, or truncated; whether material is absent relative to an ideal corner; and whether whitening is present along the cut on dark stock. This stage is deterministic image mathematics — it contains no learned model and produces identical results on identical input. Its purpose is to give every subsequent judgment an objective referent: claims that contradict the measured outline are rejected.

3.4 Multi-layer adjudication

The measured corner is evaluated by independent analysis layers: retrieval against a proprietary reference corpus of expert-verified corner examinations; a vision-language model constrained by a fixed grading rubric and a body of adjudication doctrines (see §5.3); and a proprietary trained classifier operating in shadow mode — scored daily against the ensemble but excluded from grade determination until its measured accuracy earns promotion. No single layer determines the outcome, and the physical measurement of §3.3 can veto any layer's claim. Layer weightings, thresholds, and prompts are proprietary.

3.5 Deterministic pricing

Adjudicated findings convert to a decimal through a fixed deduction structure: a defined base value for the tip's condition class, reduced in defined severity steps for whitening and for fray (Table 3). The eight corner decimals then combine through a published formula — the average of all corners, capped by the weakest corner — to produce the card's corner subgrade. The same card yields the same grade on every evaluation, and every grade decomposes into named causes.

Corner condition classes and their decimal bands are public standards:

Table 2 — Corner condition classes and decimal bands.

Table 3 — Deduction structure applied to the base condition class.

5.1 Proprietary data assets

The methodology is trained on, and continuously audited against, proprietary datasets in which every record traces to a certified, publicly verifiable graded card. Scale at the time of publication:

Table 4 — Vision™ Corner Reference Corpus, scale as of August 2026.

5.2 Continuous accuracy telemetry

Every analysis layer is scored daily over a rolling 1,000-corner window against the ensemble's final determinations. At publication, layer agreement within one grade ranges from approximately 95% (vision-language layer) to 99% (retrieval layer), with the shadow classifier near 97%. Corners where layers disagree materially are automatically queued for expert review — disagreement is treated as signal, not noise.

5.3 Expert calibration and the doctrine record

Two human governance mechanisms audit the data continuously. First, stratified samples of the reference corpus are re-reviewed by the grading lead in an assisted protocol: the system presents its grade and evidence; the reviewer confirms or corrects, with every correction requiring a written reason and annotated image coordinates, and every review stored permanently. Second, disputed grades are re-measured through a documented audit process; each confirmed dispute yields a written adjudication doctrine — a permanent rule added to the method. The method's judgment is therefore a growing body of recorded precedent rather than a static model.

The methodology's internals are proprietary; its outputs are not. Verification is a named system: the CardFax™ Registry — the permanent public record every grade is issued into. Each certificate publishes its evidence — magnified corner imagery and per-corner findings — retrievable by anyone from the certificate number, alongside the card's grade history and audit corrections. Vision™ makes the grade; CardFax™ proves it. Reproducibility is guaranteed at the system level by determinism: identical input produces identical measurement and identical pricing. External parties can therefore verify any individual grade against its published evidence without access to the method's implementation.

• Calibrated capture is required. The method does not grade from uncontrolled photographs; estimates produced from such inputs are explicitly outside this methodology.
• The shadow classifier does not yet determine grades. Our trained classifier is evaluated daily but excluded from grade determination until its measured accuracy meets promotion criteria — a deliberate governance constraint.
• Corpus density follows the graded population. The reference corpus is deepest in the 9.0–10.0 band, reflecting a deliberately curated inventory; lower bands carry fewer exemplars and correspondingly wider adjudication reliance on measurement and rubric.
• Interior damage is a separate problem. This paper covers the corner region; creases, surface wear, and edge chipping elsewhere on the card are governed by their own subgrade methodologies.

Planned extensions include promotion criteria and staged trust for the shadow classifier; deliberate corpus expansion in the sub-9.0 bands; publication of companion methodologies for centering — where every measurement is individually human-approved — and for edges and surface; and periodic revisions of this document as the doctrine record grows. Revisions will appear in the revision history below.

1. TradingCard360® Grading Standards — the published grade scale and subgrade formulas. tradingcard360.com/card-grading-software
2. "AI Card Grading: How It Works." TradingCard360® Research, 2026. tradingcard360.com/ai-card-grading
3. The CardFax™ Registry — permanent public certificate records with per-corner evidence. tradingcard360.com/card-fax-registry
4. "The CardFax™ Registry: How to Verify a Graded Card." White paper TC360-WP-2026-002, TradingCard360® Research, 2026 — the companion verification methodology. tradingcard360.com/how-to-verify-a-graded-card
5. The graded card feed — every certificate issued under this methodology. tradingcard360.com/graded-cards

Revision history.