How we get to a number you can defend.

Every price input is a completed sale. Never an asking price, never a model estimate. Ingestion runs in continuous cycles and each sale is stored with its venue, timestamp and audit trail, so any downstream number traces back to the specific listings that produced it.

Coverage is honest, not total. The first transparency report (May 16 to Jul 22, 2026) covered 65,117 completed sales totalling $25,935,941 across 2 live venues, with 18 marketplaces in the registry. Venues come online one at a time under source-policy gates. We do not scrape anything we are not allowed to.

A comp is only useful if it is the same card. Listings are matched against canonical card definitions, set, release, parallel, variant and grade, before they enter any aggregate. Ambiguous listings are excluded, never guessed into a number.

The honest gap: in that same window 12.2% of ingested listings matched a canonical definition. The remaining 87.8% were excluded from aggregates rather than force-matched. Every confidence score below reflects which subset produced it.

Fair value is reported as a median band, p25 to p75, plus a 0 to 1 confidence score. Never a single point. Confidence drops mechanically when the data is thin, stale or single-venue, so the score tells you how much to lean on the band.

Nothing is smoothed in secret. The comps behind every band are inspectable, with the suspicious ones flagged instead of quietly deleted, and a band with fewer than 3 observations in the window returns null rather than a fabricated number.

Six automated detectors run over the comp stream. Every flag is a probabilistic pattern detection carrying cited sale IDs, a time window, a confidence band and the model version that raised it, so it can be audited and disputed. Flags never name or accuse any identifiable party.

Pull-rate anomalyactive

Rare-parallel sales ratios that deviate from documented print-run distributions. Signals manufactured-scarcity inflation or selective grading.

Wash-trade repeatsbuilt

Repeated buyer/seller pairings or price patterns consistent with self-trading to establish false comps.

Shill-bid accelerationbuilt

Final-minute bid acceleration patterns inconsistent with organic demand curves.

Coordinated pumpbuilt

Cross-listing price movement correlated in time and venue beyond what market conditions explain.

Manufactured scarcitybuilt

Observed supply in the market exceeding what disclosed print runs should allow.

Consignment recyclingbuilt

The same physical card cycling through venues to refresh its comp history.

In the first report window the detectors raised 407,084 flags across 58,704 card definitions. Volume that high is the point: the filter is the review queue and the confidence band, not silence.

Every surfaced flag carries a rebuttal window. Disputes are logged append-only and reviewed. A flag dismissed under review is removed from the public surface and cannot be re-raised without fresh evidence. Card Intel is not a court of fact: dispute history is how the models get better, not how anyone gets named.