Release Log
The engine ships like software. Because it is software.
CardFax improves in versioned releases, on the record. Every entry below made grading measurably better — and none of them changed a grade already issued. The loop sharpens the future, never rewrites the past.
Releases that change how a card is measured are named for the bench instrument they turn into software — Level, Loupe, Straightedge — and each ships with its paper.
October 2026
MOSES™ 2.18: Black Backs and Chrome, Measured on Both Sides
Three corrections for black-stock and chrome cards. First, when a corner close-up shows less than half of the card against the black scan bed, the outline measurement cannot tell card from bed, so its missing-material reading is set aside and the corner is judged on tip angle, recession and roughness instead. Second, the check that recognizes a card's own uniform cut-edge line now runs on back corners: back corners had been sent to it without the back scan, so it silently never ran on any card back. Third, on chrome, refractor and Prizm fronts the check allows for the extra measurement noise reflective stock produces. Tested on six owner-audited cards: back corners that had capped grades at 8 now read 9.1 to 9.2 on their own, while corners with real, owner-confirmed damage held their low grades. Raises only on measured geometry; applies to new gradings.
September 2026
Edge Engine 3.1 and Surface: Black Stock, Read Correctly
On black-stock cards the cut edge shows its paper core as a thin bright line, and the edge examiner could call that line “very minor roughness.” The edge engine now measures each edge's line on the card's own scan: when every face carrying a mild roughness claim shows a uniform line, the claim is the stock and the edge is graded clean. Light-stock cards and any claim of chipping, whitening or fray never qualify. The surface engine now enforces in code a rule it previously stated only in words: two faint haze marks at mirror-image positions on the same side are the card's symmetric design read twice, not two matching fingerprints. On the audited card, the edges rose from 9.4 to 9.5. Applies to new gradings.
MOSES™ 2.17: Dark Corners, Read Like an Audit
On black-stock cards the examiner can read the paper core showing at the cut as rounding and whitening, and the outline measurement's missing-material reading is unreliable against a black scan bed. The engine now judges such a corner the way an owner audit does: when the tip's angle, recession, roughness and protrusion all measure square, the corner is square — the unreliable reading is set aside, not allowed to veto three good ones. A whitening or fray claim that matches the card's own uniform edge line is stepped toward it. And when a measured-square tip refutes a claim of structural wear, the corner's consensus — which that claim had overridden — is restored. On the audited card, a corner that capped the grade at 8 now grades 9.4 on its own. Raises only on measured geometry; applies to new gradings.
MOSES™ 2.16.2: Dark Print Is Not Missing Card
The corner outline measurement separates card from the black scan bed. On a dark-printed corner where the ink runs right to the cut, the print could merge with the bed and read as missing material — one audited corner measured 51% missing while its tip sat perfectly square and in place, and that phantom reading confirmed a heavy-wear grade. A square, un-receded tip cannot be missing that much of itself, so such a reading is now withheld and the corner goes to the design-aware re-examination instead. Real wear, calibrated at around 9% missing, is unaffected. Applies to new gradings and owner-authorized re-grades.
MOSES™ 2.16.1: Corner Wear Counts in Full
The corner rubric starts from the tip's shape and deducts for whitening and fray — but the result was held at the floor of the tip's band, so a sharp tip could never score below 9.7 no matter how much whitening the examiner reported. Deductions now apply in full down to the bottom of the heavy-wear band. The same audit reworded a misleading lift-check note: on dark or printed corners the check abstains because there is no bare card stock to measure, not because the crop missed the tip. Applies to new gradings; registered grades are not recalculated.
MOSES™ 2.16: Every Card Graded on Its Own Corners
A reference step could let a different, previously graded card weigh into a corner grade: when one corner closely matched a library sample, that sample's source card lent its grades to the other corners at a 60% weight. Close matches mostly reflect the same print — every copy of a card looks alike — not the same condition, so a whitened corner could be lifted toward a cleaner copy's grade. Found in an audit of a graded card whose back corner read 9.3 on its own evidence and was recorded 9.5. From 2.16 the reference card can confirm a corner but never change its grade; each card is graded on its own corners alone. Like taring a scale, the release removes a weight that was never part of the card. Applies to new gradings; registered grades are not recalculated.
Edge Engine: Memory Footprint Halved
The magnified-strip edge exam now cuts each strip directly from the scan instead of duplicating the whole card image for every strip. Peak memory fell by roughly half, which ends intermittent worker restarts that could stall an edge exam mid-batch. The strips the examiner sees are byte-for-byte identical, so no grade changes.
Edge Metrology: Every Cut Line, Traced
The engine now holds a straightedge to every edge: the cut line is traced at full resolution and compared against the line the factory intended, reporting how much of the edge is rough (coverage), how far it departs (amplitude), and what kind of rough it is — chips bitten inward versus fibers left standing by the blade. Running in shadow mode on every card graded; the published edge bands and the edge law are unchanged. On its first reading it independently found the one rough edge the examiner had charged — on both faces — and nothing anywhere else.
Read the paper: CF-WP-2026-004 — Edge Metrology →Surface: The Manufacturing-Artifact Doctrine
Press-era print dots are a signature of how a card was made, not how it was kept — so isolated minor dots now deduct linearly with a combined cap instead of snowballing under the stacking multiplier. A dot-speckled but otherwise clean junk-wax surface lands in the mid-8s, not the 5s; genuine damage still stacks. Like a loupe in a grader's hand, the release is about telling what a mark is before deciding what it costs.
Read the paper: CF-WP-2026-003 — Surface Defect Deductions →MOSES™ 2.15: Corner Lift Detection
A grader's hands caught a slight corner curl the scan had read as sharp. The engine now measures the shadow a lifted corner casts on the scan bed — normalized against the card's own eight corners and confirmed against the front/back mirror pair — on every card graded, and mirror-confirmed curls deduct at grade time. Shipped in shadow mode and activated the same day after the flag went two-for-two against physical review. A level detects the one thing the eye reads wrong: out-of-plane.
Read the paper: CF-FR-2026-001 — The Corner That Graded Sharp →Batch Speed Grading
A graded seed card now teaches its copies: centering runs hands-free across duplicates of the same card, and miscuts are automatically parked for individual review instead of slowing the batch.
Nightly Population Identity Audit
Every population count is now backed by a nightly registry-wide identity audit — player, set, number, and variation are verified before a card is counted, so one card can never inflate two population lines.
Blind Corner Calibration Program
Examiners review corners in the 9.0-10.0 band without seeing the machine's answer. Every agreement confirms the model; every disagreement becomes training signal for the next MOSES™ release.
August 2026
Measured Evidence on the Public CardFax
The geometry behind each corner grade — the silhouette measurements themselves — is now published on the public CardFax record, so anyone can see why a card earned its grade.
MOSES™: Silhouette Geometry Layer
A new fourth analysis layer measures the outline of every corner directly, giving MOSES™ a pure-geometry check on every grade the ensemble produces.
Surface Back Parity
Card backs now receive the full surface treatment — dedicated speck detection and magnified axis review — matching the scrutiny fronts have always had.
Speed Grading
The grading pipeline was rebuilt end to end, cutting average grading time from 223 seconds to 58 seconds with zero change to grading standards.
Set-Aware Centering Priors
Centering is now guided by priors for the specific set being graded — the engine knows how wide a set's borders run before it measures your copy.
July 2026
Automatic Card Detection at Intake
The scanner now detects, crops, and fits the card automatically at capture, anchoring every downstream measurement to the card itself rather than the photo.
Full-Pipeline Performance Overhaul
Bounded image decoding and modern formats across the pipeline — large scans no longer trade speed for measurement quality.
June 2026
MOSES™ Four-Layer Ensemble in Production
The complete corner ensemble — detection, library similarity, AI examiner, and rules — verified end to end in production, with the training pipeline refreshing the sample library every 30 minutes.
Every release starts in the loop
Releases are not roadmap promises — they come out of the closed loop of grading, recording, and calibration that runs every day.
How the Engine WorksFrequently Asked Questions
What are CardFax engine releases?
Versioned changes to the grading engine, published like software release notes: what changed, why, and what it measured before and after. Grading behavior never changes silently.
Do engine updates change existing grades?
No. Formula and calibration changes apply to new gradings only. Registered grades are permanent records of the standard they were graded under.
Why publish engine changes at all?
Because auditability is the product. A grading system that changes quietly cannot be trusted with a market; one that publishes every change can be checked by anyone.
How often does the engine improve?
Continuously — the training loop refreshes on production measurements around the clock, and named releases ship when behavior changes enough to document.
