Verified PSA-Protocol // Logic Engine

PRECISION
BY LOGIC.

Should you send this card to PSA? A submission ties up $25–$600 in fees and up to 7 months of capital lock. Only 43% of submissions return a Gem Mint 10. AuraGrade tells you which side of that line your card is on — before you ship.

Inspection Depth
47 Points / Scan
Ground Truth
PSA Cert-Verified
Centering // Front
52 : 48 [PASS 55/45]
SOBEL PIXEL COUNT
Micro-Zoom // Corner
R: 12.4px [IDEAL]
FRAY_PROB: 0.02
Perimeter Trace
0.2mm SILVERING
CHIP_PROB: 0.03
Spectral // Surface
0 DIMPLES // 0 PRINT_LINES
STATE: READY
INFERENCE_ID: 18884B69 PRECISION: ±0.3px
YOLO_CONF 0.00
INSPECTION_PTS 0 / 47
MTL_GRADE_DISTRIBUTION — awaiting inference —
PSA 10
PSA 9
PSA 8
Capture Protocol // Edge Validation

Garbage in, garbage out. Never.

A client-side validator gates every upload at the edge before any inference cycle is spent. Sub-threshold captures are rejected with a single actionable retake instruction. Image or video — the engine accepts both.

Required Captures PER SCAN

Tap any capture to preview it in the validator →

Quality Thresholds

Evaluated client-side via WebGL/HTML5 Canvas. No network round-trip required to reject a bad capture.

Min Resolution
Card must occupy ≥60% of frame
1500 × 2100 px
Sharpness (Laplacian Var.)
Rejects motion blur and misfocus
≥ 250
Exposure Range
No clipped histogram tails
0.05 – 0.95
Specular Glare
Hotspot as % of card area
< 4%
Keystone Angle
Max tilt before homography rejects
< 8°
Corner Visibility
YOLO must locate all four anatomical corners
4 / 4

Edge Validator

Live // 60 Hz

Retake Feedback

Failed captures return one actionable instruction — never a generic error. Examples returned by the validator:

RETAKE Top-right corner cut off. Reframe so all 4 corners are visible inside the guide.
RETAKE Specular glare on 11% of card surface. Tilt the light source ~30° off-axis or switch to diffuse lighting.
RETAKE Sharpness 142 (need ≥250). Hold device steady; tap to refocus on the card center.
RETAKE Keystone tilt 12.4°. Hold the camera parallel to the card surface — keep the lens directly above.
RETARGET Back of card missing. Flip the card and capture the reverse side to complete the scan.
RETARGET Bottom-left corner needs a macro. Switch to corner-zoom mode for high-stakes pre-grading.
Video Mode // Multi-Frame Inference

Record an 8-second walkaround. Skip the retake loop entirely.

A slow rotational pass captures every angle in one motion. Our frame-extraction pipeline scores each frame on sharpness, glare, and perpendicularity — then picks the optimal frame per pillar. Multi-frame fusion cancels specular reflections that defeat single-image captures, which is what makes surface-defect recall so much stronger here than on competitor platforms.

~240
Frames Scored
4–6
Best Frames Fused
0
Retakes Required
WALKAROUND · 08.2s 240 fps · scoring
F1
41
F2
F3
62
F4
F5
F6
46
0.0s
8.2s
✓ 3 BEST FRAMES FUSED SPECULAR CANCELLED · 0 RETAKES

The Four Pillars of Grade

A ConvNeXtV2 backbone with separate Multi-Task Learning heads for each pillar. Each pillar is evaluated independently, then the Score Aggregator enforces PSA's hard-cap logic before a final distribution is returned.

01

Centering

Algorithmic pixel counting via Sobel edge detection — not a deep-learning guess. Enforces PSA's 55/45 front threshold for Gem Mint 10 with mathematical certainty.

52 : 48
L:8.4px / R:7.7px
Method: Sobel + Morph. Ops
02

Corners

Dedicated MTL head trained on 20K hard-negative crops. Detects micro-fraying and whitening at radii invisible to the naked eye.

10× // R:12.4px
Method: MTL Head + MSE Loss
03

Edges

Perimeter segmentation identifies silvering and rough factory cuts on vintage stock — defect classes most consumer-grade graders miss entirely.

0.2mm SILVERING
Perimeter Trace
Method: Perimeter Segmentation
04

Surface

Trained against synthetic glare and shadow augmentations so the network learns invariant defect features — dimples, print lines, scratches — even under bad smartphone lighting.

INDENT_MAP: 0 DEFECTS
Spectral 8×4
Method: MTL + Glare Augment.
Grading Bands // What PSA Is Looking For

The 3-grade chasm your submission is rolling against.

The final grade is the floor of all four pillars — one bad corner caps the whole card. Only 43% of PSA submissions return a Gem 10. The valuation gap between a 10 and a 9 on the same card routinely runs 3× to 10×. AuraGrade tells you which band you're holding.

PSA 10
Gem Mint
43%
Gem Rate
FLAWLESS
Centering55/45 or better
CornersSharp at 10× mag
EdgesFlawless to eye
SurfaceZero defects
Base Set Charizard $5,000–8,000
1st Ed Shadowless $300k–550k
Decision
Submit Immediately
PSA 9
Mint
~30%
Next Band
1 MICRO-FRAY
NEAR-PERFECT
Centering60/40 or better
CornersMicroscopic fray OK
EdgesMinor whitening
Surface1–2 minor marks
Base Set Charizard $1,500–2,500
1st Ed Shadowless $30k–50k
Decision
Worth The Gamble?
Let AuraGrade quantify the upside
PSA 8
NM–Mint
~15%
Below Mint
MULTI-DEFECT
EDGE WHITENING
Centering65/35 or better
CornersSlight fraying
EdgesNoticeable whtnng
SurfaceVisible defects
Base Set Charizard $700–1,200
1st Ed Shadowless $8k–15k
Decision
Don't Submit
Submission fee likely exceeds grade lift
The Floor Rule
Final PSA grade = min(Centering, Corners, Edges, Surface). One micro-frayed corner caps the whole card. AuraGrade's MTL aggregator mirrors this exact hard-cap logic — so the predicted grade reflects PSA's actual grading psychology, not an averaged guess.

A wrong submission
is months of locked capital.

PSA charges $25–$600 per card and ties up your asset for 5 business days (Super Express) to 7+ months (Value tier). 57% of submissions return below Gem Mint 10. The expected value of submitting blind is brutal — AuraGrade screens the decision in seconds for the cost of a pack of gum.

PSA Tier Pricing // Feb 2026
Value Bulk$25 · 150+ days
Economy$50 · ~65 days
Regular$80
Express$200
Super Express$400 · 5 days
Walk-Through$600 · same-day
57%
Of submissions return
below Gem Mint 10
$0.33
AuraGrade pre-screen
per card // sub-10s

Avoiding even a single mis-submission pays for years of Pro tier.

Subscription Tiers
Basic
$0
· 3 free scans per day
· Standard AI grade prediction
· Algorithmic centering report
Recommended
Pro
$9.99/mo
· Unlimited scans
· Full 47-point inspection report
· High-resolution PDF certificates
· Priority inference queue
· Market value estimation
Dealer
$49/mo
· Everything in Pro
· Bulk upload (batch > 100 cards)
· REST API access
· CSV export for inventory systems
Competitive Intelligence

The axes that actually matter.

First-generation AI graders validated the demand but failed the execution. We compete on the dimensions where their architectures collapse: surface-defect recall under variable lighting, sub-grade granularity, cryptographic transparency, and permanent data ownership.

CardGrader.ai
$5-$75/mo // credits
Most sophisticated incumbent. Strong base-rate accuracy. Opaque outputs, restrictive credit system, no audit trail.
CardBoss
UX broken
Reports cannot be regenerated after navigation. Market estimates inaccurate by 3-5x. No usable historical inventory.
12K
eBay // PSA-Verified
5K
Hard-Negatives
3K
User Telemetry
Training Corpus // v3.0

A 20,000-image defensible moat.

Each label is cryptographically verified: OCR extracts the PSA cert number from the slab, then queries PSA's official Verification API to confirm the grade, set, year, and qualifiers. No seller descriptions. No mislabeled samples. No training contamination.

eBay // PSA-Verified Scrapes
12,000 sold-listing photos with cryptographically verified PSA labels via the official Cert API. Anchors the model to real-world market distribution.
Synthetic Hard-Negatives
5,000 parametrically generated cards via Albumentations: synthetic glare, micro-scratches, contrast/brightness shifts. Forces invariant defect representations and closes the surface-detection blindspot.
Live User Telemetry
3,000 freemium-beta captures with downstream PSA verification. Closes the loop between our prediction and the slab's eventual ground-truth outcome.
47 pts
Inspection Depth
±0.5
Grade Margin
A100
Modal GPU Cluster
Three-Phase Roadmap

From pre-grader to grading authority.

The pre-grading product is the wedge. The proprietary corpus is the moat. The endgame is replacing the incumbent grading paradigm with cryptographic transparency.

RAW SCAN
PSA 10: 78%
Phase 01 // Months 0–9

AI Pre-Grader

SaaS pre-submission filter. Every paid scan generates cash flow and crowdsources real-world user images into the proprietary training corpus.

ConstraintAccuracy / Depth
RevenueSubscription
OutputProbability Dist.
AURAGRADE PSA 10
ENCAPSULATED
Phase 02 // Months 9–24

Hybrid Encapsulation

High-confidence pre-scans get invited to mail in for physical processing under controlled multi-spectral lighting. AI does 95% of evaluation; human technicians do final QA + sealing.

ConstraintPhysical Throughput
RevenueEncapsulation Fees
OutputTamper-Evident Slabs
AURAGRADE 10 GEM
AG-4F8A2C
Defect Map →
Cryptographic
Phase 03 // Year 2+

Cryptographic Authority

Every slab carries a QR linking to a permanent on-chain ledger entry with the pixel-level defect map and per-pillar reasoning. Algorithmic transparency replaces the opaque legacy integer.

ConstraintMarket Adoption
RevenueAuthority Fees
OutputAuditable Defect Map