How AgentGrail AI Grades Cards: Rubric, Confidence, and Honest Limits

Most AI grading tools give you a number. We give you a verdict, a confidence level, and — when we genuinely can't tell — an honest "we can't grade this photo."

That's not a limitation we're apologizing for. It's the design. Here's exactly what our AI measures, how it measures it, where it excels, and where it stops.

The Four Grading Criteria

Every PSA, BGS, SGC, and CGC grade is built on the same four criteria. Our AI evaluates the same rubric the professional graders use:

Criterion	What It Covers	AI Reliability
Centering	Border width ratios — left/right and top/bottom	High
Corners	Sharpness and wear at all four corners	Good
Edges	Nicks, chipping, and condition along all four edges	Good
Surface	Scratches, print lines, haze, and surface wear	N/A — see below

Centering

What it is: The ratio of the border widths on each axis. PSA 10 requires no worse than approximately 55/45 left-to-right and top-to-bottom; 60/40 is the cutoff for PSA 9.

What we measure: Geometric border ratios from the card image. Centering is a pixel-level geometric property — it doesn't require physical handling. This is one of the most reliable things AI can measure, and the numbers are real: 54/46, not "looks good."

Known limit: If the card is at an angle, a border is cut off, or the background blends with the card edge, we'll flag the measurement as unreliable rather than invent a ratio.

See also: how professional graders evaluate each criterion.

Corners

What it is: The condition of all four corners — sharp and square for Gem Mint, with progressive deductions for soft corners, fraying, and obvious wear.

What we detect: Edge geometry and texture changes at each corner point. Our model identifies the visual signatures of wear — rounding, tip fraying, and disruption at the corner itself.

Best results: Close, even lighting with all four corners in frame. If a corner is partially obscured by a sleeve or shadow, we'll express uncertainty on that corner rather than assume it's clean.

Edges

What it is: The condition along all four edges — chipping, nicks, and indentation visible as dark marks or irregularities along the edge line.

What we detect: Edge sharpness and disruption patterns. Edge defects show up clearly in properly lit photos. Heavy wear (obvious chipping) is caught reliably; very minor edge roughness at PSA 9/10 boundaries is where we're more likely to express a REVIEW verdict rather than commit.

Surface: Why We're Honest That We Can't Grade It

Surface is the one criterion we do not grade — and we tell you that directly instead of fabricating a score.

Here's why:

Surface defects — scratches, print lines, print imperfections, haze, and roller marks — are often invisible in head-on photography. They become visible under raking light: light angled low across the card surface, which casts shadows from even minor surface disruptions. This is why every professional grader's first move is to tilt the card under a lamp.

A head-on phone photo of a card, even a technically excellent one, cannot reveal most surface issues. The light hits the surface flat, and the defects are invisible. An AI trained on such photos produces surface scores that appear confident — but reflect camera artifacts, not actual card condition.

We made a deliberate choice: if we can't measure something reliably, we say so rather than fabricate certainty.

What this means in practice:

• Our verdict weighs centering, corners, and edges — the criteria visible in standard photos
• Surface returns an honest N/A rather than a guess
• For premium cards where surface matters most (vintage, high-pop PSA 10 candidates), physical inspection under raking light is irreplaceable — and we'll tell you that

A grader who says "I can't see the surface in this photo" is more valuable than one who invents a surface score. That's the standard we hold ourselves to.

The Verdict System: BUY / PASS / REVIEW

Our output isn't a raw numeric prediction — it's a verdict with calibrated confidence.

BUY

The card shows strong indicators for Gem Mint or near-Mint condition across the criteria we can evaluate. We set confidence thresholds high before issuing a BUY — we'd rather miss a BUY on a borderline card than send you to PSA with a card that grades out as a 7 or 8.

PASS

Visible defects detected that are likely to reduce the grade below the threshold where submission makes financial sense. The issues are clear enough that grading is unlikely to return a premium grade.

REVIEW

We saw something, but we're not confident enough in either direction. Common reasons:

• Centering is borderline (within a few percent of the PSA 9/10 threshold)
• Minor corner or edge wear that's hard to classify at the grade boundary
• Photo quality limits what we can measure reliably
• The card has characteristics that warrant closer physical inspection

REVIEW is not a soft PASS. It means "look closer before deciding." Some REVIEW cards are excellent; some should be passed. A human look — ideally under raking light — is what the REVIEW verdict is asking for.

See also: the grade premium difference between PSA 10 and PSA 9.

How Confidence Works

Every verdict comes with a confidence score. We apply calibrated thresholds before issuing BUY verdicts: the AI needs to clear a higher bar for Gem Mint candidates than for strong-but-not-perfect grades. Below the threshold, the verdict falls back to REVIEW.

This calibration reflects real-world grading outcomes, not training accuracy on a benchmark. We've tuned the thresholds to minimize expensive false positives — cards that look BUY to the model but grade out at PSA 7 or 8.

You can also adjust your own threshold using the pickiness slider in your account settings. Stricter settings require higher confidence before a BUY verdict; more lenient settings cast a wider net. The default is calibrated for collectors who want to minimize bad submissions.

What We Cannot Do (And What This Means)

Capability	AI Status	What to Use Instead
Centering measurement	Reliable	—
Corner condition (clear photos)	Reliable	—
Edge condition (clear photos)	Reliable	—
Surface grading	Not available — honest abstention	Physical inspection under raking light
Trimming / alteration detection	Low confidence — not recommended for authentication	PSA authentication service, physical measurement
Refractor / foil parallel effects	Partial — light-angle effects not always visible	Title/description + physical check
Very small or blurry photos	We'll tell you the photo is insufficient	Retake with the tips below

How to Get the Best Results

Our analysis quality is bounded by photo quality. The same card can produce a REVIEW verdict in a poor photo and a confident BUY in a proper one. Here's what makes the difference:

1. Shoot directly overhead, card flat. Any angle introduces border distortion that throws off centering calculations.
2. Fill the frame. All four borders should be visible, but the card should fill most of the image.
3. Even, diffused lighting. Harsh directional light creates shadows in corners and edges that look like defects. A lightbox or diffused LED panel is ideal.
4. Submit both sides. Our dual-scan mode weighs front and back for a more complete picture. The back often shows centering and surface issues the front doesn't reveal.
5. Remove penny sleeves for the scan. Plastic catches light in ways that obscure corners. Grade raw or in a top loader with the card visible.

See also: how to photograph sports cards for the best scan results.

The Honest Bottom Line

AI card grading is a pre-screening tool. It replaces the first pass of sorting through a stack — the obvious passes, the clear BUYs, the borderline cases worth closer attention. It does not replace a human grader with raking light and a loupe.

What it does well — centering measurement, detecting obvious corner and edge defects, calibrated confidence that tells you how much to trust the verdict — it does at a pace and scale no human can match.

What it can't do, we'll tell you. "Can't grade the surface from this photo" is not a failure message. It's the correct answer.

---

Frequently Asked Questions

How accurate is AI card grading compared to PSA?

AI grading is accurate for the criteria visible in standard photos — centering, corner condition, and edge condition — but it is not a replacement for professional grading. AgentGrail is designed as a pre-screening tool: it catches obvious passes, surfaces clear BUY candidates, and flags borderline cases before you spend money on a submission. Surface defects, which require raking light to detect, are outside what any AI can reliably assess from a flat photo.

What does a REVIEW verdict mean in AgentGrail?

REVIEW means the AI detected something worth closer inspection but is not confident enough to commit to BUY or PASS. Common triggers include centering within a few percentage points of the PSA 9/10 threshold, minor corner or edge wear at a grade boundary, or photo quality that limits what can be measured reliably. REVIEW is not a soft PASS — some REVIEW cards are excellent candidates once examined in hand under raking light.

Why doesn't AgentGrail grade card surface condition?

Surface defects — scratches, print lines, haze, and roller marks — are only visible under raking light, where a low-angle beam casts shadows across even minor surface disruptions. Head-on photography, including technically excellent phone photos, cannot reveal these defects. Rather than produce a surface score that looks confident but reflects camera artifacts instead of actual card condition, AgentGrail returns an honest N/A and tells you to inspect the surface in hand.

What confidence level is required before AgentGrail issues a BUY verdict?

AgentGrail applies calibrated confidence thresholds before issuing a BUY — the bar for Gem Mint candidates is higher than for strong-but-not-perfect grades. Below the threshold, the verdict falls back to REVIEW. You can adjust your own threshold using the pickiness slider in your account settings; the default is tuned to minimize expensive false positives, meaning cards that appear to be BUYs but grade out at PSA 7 or 8.

Does submitting both the front and back of a card improve grading accuracy?

Yes. AgentGrail's dual-scan mode weighs the front image at 70% and the back at 30% to produce a combined verdict. The back of a card often reveals centering issues and corner wear that the front obscures, and it provides independent confirmation of edge condition. For any card you are seriously considering submitting to PSA, scanning both sides gives the AI more signal and reduces the chance of a verdict based on a single misleading angle.

Can AgentGrail detect if a card has been trimmed or altered?

Trimming and alteration detection is outside AgentGrail's reliable capability, and we do not surface it as a grading signal. Detecting trimming requires precise physical measurement — calipers and comparison to factory-standard dimensions — along with examination under magnification that no standard photo can replicate. For authentication concerns, PSA's authentication service and physical measurement with calipers are the appropriate tools.