Our Methodology
AgentGrail is a pre-grading tool, not a certified grading service. We analyze photos to help you decide whether a card is worth sending to PSA, BGS, or SGC — before you spend $30–$100+ on a submission.
Below: how we identify, price, and AI-grade cards — and where our system has honest, known limitations.
Last updated: June 2026 · See changelog
Donnie Laur
Founder, AgentGrail
1. Card Identity
Every card in the AgentGrail catalog goes through an identity waterfall — a sequence of progressively more expensive lookups that stops as soon as a high-confidence identity is established.
1a. PSA Certificate Lookup
For graded slabs, we first check the PSA public API against the slab's certification number. PSA cert data has the highest authority (confidence 1.0) because it was directly recorded at the time of grading. Fields: player, year, set, card number, variant, grade.
1b. CardSight Catalog
For cards without a cert number, we check CardSight's ~8 million card catalog. CardSight returns structured identity data (player, year, brand, set, card number) at sub-second latency.
1c. Claude Haiku Vision
When CardSight misses or returns low-confidence data, we send the card image to Claude Haiku (an Anthropic AI model) with a structured prompt asking for identity fields. Haiku has broad knowledge of modern card sets and performs well on standard trading card photos.
1d. eBay Listing Title Parse
As a final fallback, we parse the eBay listing title using a custom sports-card-aware parser. This is the least reliable method; cards identified only via title parsing are marked lower-confidence in the catalog.
2. Pricing Data
Prices on AgentGrail reflect recent real-world transaction data, not ask prices.
eBay Sold Listings: Our primary pricing signal. We query the eBay Browse API for recently-sold comparable listings (same player, set, year, grade tier) and compute a median price from the most recent transactions. Prices update daily via our price-refresh job.
PriceCharting: For TCG cards (Pokémon, Magic: The Gathering, Yu-Gi-Oh!, One Piece), we supplement eBay data with the PriceCharting catalog, which tracks long-run price history for singles. TCG prices on AgentGrail use a weighted blend of both sources.
Currency & freshness: All prices are in USD. Each card page displays the most recent price snapshot date. Prices that are more than 30 days old are shown with a staleness indicator.
3. AI Grading
The AgentGrail AI grading system predicts whether a card is a likely PSA 10 (Gem Mint) candidate — the industry's top grade — and outputs a BUY, PASS, or REVIEW recommendation.
Sub-criteria extraction
We score four grading criteria on a 1–10 scale: centering (left/right and top/bottom), corner sharpness, edge wear, and surface condition. These scores feed into the grading head.
CLIP visual embedding
A 512-dimensional visual embedding is extracted using OpenAI's CLIP ViT-B/32 model. This captures global visual quality signals that sub-criteria scores can't fully quantify.
Grail Head classifier
A neural-network classifier ("Grail Head") takes the 512-dimensional CLIP image embedding and outputs the BUY/PASS verdict probability. This is the primary grader. The 4 sub-criteria scores (centering / corners / edges / surface) are shown on your scorecard for explainability — they are not inputs to this verdict, so an unmeasurable sub-criterion never silently moves your BUY/PASS result.
Confidence gates
A prediction is only labelled BUY when the Grail Head output exceeds the user's configured confidence threshold (default 0.65 for B-grade, 0.75 for Gem Mint). Predictions below the gate are shown as REVIEW.
4. What BUY / PASS / REVIEW Means
BUY
Strong Gem Mint / Mint (PSA 9-10 band) candidacy based on centering, corners, and edges visible in the photo. We deliberately call the 9-10 band, not a single grade — that line is too noisy to promise from photos, and surface defects invisible to the camera could still drop the grade. It does not factor in current raw prices, grading fees, or market demand unless you use the Flip Calculator.
PASS
We see enough evidence of defects to expect a grade of PSA 9 or lower. Submitting a PASS card will likely cost more in grading fees than the grade improvement is worth at current market prices.
REVIEW
The photo doesn't give us enough information for a confident call. This is an honest “I don't know” — not a failure. Take a better photo (remove from sleeve, plain dark background, fill the frame), or just submit and let the grader decide. Trust REVIEW: we say it because the photo is ambiguous, not because the card is bad.
The strictness setting (Conservative / Balanced / Aggressive) shifts how much confidence we require before calling BUY. Conservative requires near-certainty; Aggressive surfaces anything that could be profitable after grading fees.
5. Data Sources & Licensing
| Source | What we use it for | License |
|---|---|---|
| PSA | Cert data, population reports, graded-card identity | PSA Data License Agreement (signed 2026-05-13, §8 attribution required) |
| eBay | Search, sold listing prices, card images | eBay API Terms of Use — listed prices displayed as live quotes with rel=sponsored on affiliate links |
| CardSight | Card identity (player, set, year, number) | CardSight API License (paid tier) |
| PriceCharting | TCG card price history (Pokémon, Magic, Yu-Gi-Oh!, One Piece) | PriceCharting data via daily CSV export |
| Scryfall | Magic: The Gathering card catalog (names, sets, oracle text) | Scryfall data is provided under Creative Commons Attribution 4.0 |
| Pokémon TCG API | Pokémon card catalog (sets, card numbers, rarity, images) | Pokémon TCG API (free, attribution required) |
| Anthropic Claude | Card identity vision fallback (Haiku), grading sub-criteria (Sonnet) | Anthropic API — predictions are AgentGrail AI outputs |
6. Known Limitations
⚠️ Surface defects require specialized lighting
Our AI cannot reliably assess surface scratches, print defects, or print lines from standard eBay listing photos. Surface grading requires raking light photography (angled light source) to reveal low-relief defects. Standard head-on photos physically hide these defects. We disable surface grading in our public-facing score and exclude surface sub-criteria from the final grade.
⚠️ Minority-sport classifier reliability
Our sport classifier performs well on the highest-volume categories (baseball, basketball, football, Pokémon). For lower-volume sports (soccer, hockey, racing, tennis, golf, fighting, wrestling, MMA), the model is less reliable due to limited training data, so we fall back to title parsing rather than visual classification.
⚠️ Parallel/refractor detection is title-first
Refractors, prizms, shimmers, and other parallel variants only appear under specific lighting conditions. We extract parallel type from the eBay listing title (e.g. 'Prizm Green Wave Refractor') rather than from the image. Visual detection of parallels from standard photos is unreliable and is not part of our grading signal.
⚠️ Price data freshness
Prices are refreshed daily via the eBay Browse API. Cards with low search volume may have price data that is 1–7 days old. Prices shown are medians from recent sold listings, not ask prices. Do not rely on these prices as financial advice.
⚠️ Full-art / borderless cards report centering as N/A
Centering is computed from inner border widths. Full-art, borderless, and super-short-print cards with no visible border cannot be measured — we return an honest N/A rather than a fabricated number. A centering N/A does not affect the BUY/PASS verdict, which is produced from the CLIP image embedding — the per-criterion scores are shown for explanation, not used as verdict inputs.
⚠️ Photo quality degrades all measurements
Glare, blur, off-angle shots, and cards in sleeves, cases, or holders degrade centering, corner, and edge measurements. When photo quality makes a confident assessment impossible, we return REVIEW and tell you why. Removing the card from its sleeve and shooting straight-on under even lighting produces the most accurate results.
⚠️ AI-generated predictions may be inaccurate
All grading predictions, BUY/PASS/REVIEW recommendations, and valuation estimates are generated by AgentGrail AI and may be inaccurate. Always verify a card's condition and market value independently before making a purchase or grading decision.
7. Accuracy — What We Say and Don't Say
We don't publish a single accuracy percentage because it depends heavily on photo quality and card type. What we do: run regular benchmark evals against known-graded cards and retrain when accuracy drops. What we don't do: claim accuracy we can't verify, or pretend phone-photo AI can do what raking light and a trained human eye do in a grading room.
When our model returns REVIEW — trust it. It means the photo is ambiguous. Take a cleaner photo or submit and let the grader decide. A REVIEW is an honest “I don't know,” not a wrong answer.
8. Feedback & Corrections
The AgentGrail AI improves through user corrections. When you mark a prediction as incorrect on the analysis panel, that signal is weighted by confidence and incorporated into the next model retraining cycle.
To report a data error (wrong player name, incorrect year, missing card from the catalog):
- Email [email protected]
- Use the feedback button on any card page
- Open a GitHub issue on the public roadmap
PSA data corrections should be submitted directly to PSA at psacard.com. AgentGrail reflects PSA's official data and cannot override cert records.
Tools Built on This Methodology
- Free Centering Tool — enter border measurements, get PSA / BGS / SGC max grade ceiling. No signup required.
- Flip Calculator — model grading ROI with real PSA 10 / 9 prices and submission fees.
- AI Search — scan eBay listings in real time and get BUY / PASS / REVIEW verdict with per-criterion breakdown.
Questions about our methodology? We're transparent about how our system works and welcome feedback from the collector community.
Try AI card grading free →