Is AI Card Grading Accurate?
An AI grade is only as good as its method and the photographs behind it, so the way to judge one is to check whether both are published. DCM Optic evaluates every card three independent times and takes the median as the grade, scores it against a rubric published at /grading-standard, publishes its known limitations, attaches an image confidence letter from A to D with an uncertainty range, and makes every grade publicly verifiable by serial.
Updated August 2026. This page describes what DCM Optic does. It makes no claim about how a human grader at another company would grade the same card.
How DCM Optic produces a grade
Six steps, in order. None of them is a black box you have to take on faith, because the rubric they score against and the limits they operate under are both published.
Three independent evaluation passes
Every card is evaluated three separate times against the same published rubric. The passes do not see each other. A defect has to survive more than one pass to carry weight, which is what stops a single misread from setting the grade.
Median consensus, server-side
The final grade is the median of the three passes, computed on our servers rather than chosen by the model. That makes the result reproducible arithmetic on top of the three reads rather than a fourth opinion about them.
Magnified crop inspection
Corners, edges and surface zones are re-examined on magnified crops rather than on the whole-card image alone, because a corner is a few dozen pixels in a full-card photo and a few thousand in a crop.
Weakest link, four subgrades
Centering, corners, edges and surface are scored separately, and the final whole-number grade is the lowest of the four. Structural damage can cap the grade independently. Nothing is averaged upward.
Image confidence A to D, with a range
Every grade carries a letter describing how well the photographs supported the read, plus an uncertainty range on the grade itself. Publishing the doubt is the part most grading outputs leave off.
Published rubric, published limits, public record
The rubric is at /grading-standard and the limitations are at /grading-limitations. Every grade is verifiable at /verify by serial, and the platform-wide grade distribution is open at /pop.
The consistency argument
DCM Optic grades card #1 and card #4,000 identically. It does not have a bad day, it does not know your card is valuable, it has no submission tier and no name on the flip to react to, and it has no reason to nudge a 10 down to a 9. There is no fatigue at the end of a long session and no order effect from whatever it graded a minute earlier.
That is a claim about consistency, which is measurable, rather than a claim about being right more often than someone else, which we do not make.
Some things genuinely do not photograph: a crease that only catches the light on a tilt, a soft corner you feel more than see. At submission you can flag what the camera cannot resolve, and the analysis takes that input. Machine consistency where consistency matters, human context where it does not exist yet.
Where a photo-based grade stops
Grading from photographs is not perfect, and any service claiming otherwise is selling you something. Here is the honest list.
- Photo quality is a gate, not a detail. Glare on foil and gloss, soft focus, low resolution and a bad crop all reduce what can be resolved. That is what the A-to-D confidence letter measures, and a C or a D means retake before you act on the number.
- Autograph authenticity is not graded. A photo cannot authenticate a signature. Autographs are designated rather than authenticated, and authentication is a job for a service that does it.
- Alteration and trimming. Detecting a trimmed edge or a restored corner often needs the card in hand, measured and viewed under raking light. A photo-based grade does not stand in for that.
- What the camera cannot see. Some defects only appear on a tilt. Flag them at submission and the analysis takes the input.
- It is not a sealed slab. A DCM grade is a documented evaluation and a printable label for a holder you own. It is not a slab from a mail-in grader and it is not registry-eligible.
The full version of this list is maintained at /grading-limitations.
Check it yourself
The rubric is published
The versioned standard every pass scores against is at /grading-standard. You can read the criteria before you grade anything.
Every grade is verifiable
Each slab label carries a QR resolving to a public page at /verify/[serial] showing the grade, the subgrades and the reasoning. A buyer never has to take your word for it.
The distribution is open
The full grade distribution across every public card graded on the platform is at /pop, broken out by category. If DCM handed out 10s freely, that page would show it.
The report shows its work
What comes back is documented at /reports-and-labels: the subgrades, the defect log and the confidence letter, so you can audit the grade rather than accept it.
Most of your collection ends here. The top slice still goes out — we’ll tell you which.
Frequently asked questions
Is AI card grading accurate?
An AI grade is only as good as its method and the photographs it is given, which is why DCM publishes both. Every card is evaluated three independent times and the median becomes the grade, so a single outlier read cannot decide the result. The rubric those passes score against is published at /grading-standard, the known limitations are published at /grading-limitations, and every grade carries an image confidence letter from A to D plus an uncertainty range. DCM does not claim to predict how a human grader at another company would grade the same card.
Can AI card grading be trusted?
Trust in a grade comes from being able to check it, so DCM is built to be checked. Every deduction is logged with a written reason naming what was seen and where, every grade is publicly verifiable at /verify with a serial number, and the full grade distribution across the platform is published in an open population report at /pop. If a grade looks wrong to you, the report tells you exactly which subgrade caused it and why.
Is AI card grading biased?
DCM Optic has no information that would let it favour one card over another. It does not know the card’s market value, it does not know who submitted it, it has no submission tier, and it grades card number one and card number four thousand the same way. There is no fatigue at the end of a long queue, no reputation attached to a name on the flip, and no order effect from what it graded before. That consistency is the concrete thing an automated grader offers.
How does DCM Optic decide a grade?
It scores four subgrades and the final grade is the lowest of them, a weakest-link rule. Centering, corners, edges and surface are each scored on the published rubric, and the card cannot grade higher than its worst area. Grades are whole numbers from 1 to 10. Because the rule is public, you can reproduce the arithmetic yourself from the report.
What can AI grading not do?
It cannot judge anything the photographs do not contain, and it does not authenticate. Autograph authenticity, alteration and trimming detection, and anything that needs the card in hand under raking light are outside what a photo-based grade covers, and DCM designates rather than opines on them. Glare on foil, soft focus and low resolution all genuinely reduce certainty, which is what the A-to-D image confidence letter is for.
Why does my grade come with a confidence letter?
Because a number without a confidence level is a guess wearing a suit. The letter grades the photographs, not the card: an A means the images resolved everything the rubric needs, a C or D means glare, blur or crop cost the evaluation certainty. A low letter is an instruction to retake the photos before you act on the grade, not a verdict on the card.
Does DCM grade the same card the same way twice?
That is the design goal, and the three-pass median exists to make it true. Running three independent evaluations and taking the median damps the variance any single pass would introduce, and the report publishes how much the three passes agreed. Where they disagree materially, the uncertainty range widens so you can see it.
Read the rubric, then test it on your own card
Two free grades to start. Four subgrades, a written reason for every deduction, and a confidence letter that tells you when the photos let it down.
Grade Your First Card Free