We track 24,200 PSA registry lines and publish a population figure for 19,950 of them. The other 4,250 are measured, stored and deliberately not shown. This is the article about those, because a number withheld on purpose is a different thing from a number we do not have, and the difference is most of what separates a research product from a scraper.
Why a population can go down
A PSA population is a lifetime cumulative count. Within one registry line it can only rise, because graded slabs are not ungraded. So a reading that falls is not a correction, it is proof that the two readings came from two different lines.
Base Set Venusaur is the clean case. We read it at 311 graded copies on August 8 and watched it inch to 314 by September 8, which is what a real registry does. On September 15 the same product id came back at 46,802, a 149-fold jump in a week. Base Set Mewtwo did the same thing on the same day, 363 to 31,155. Neither card was suddenly graded thirty thousand times. The provider re-pointed the product id at a different registry line, and everything before the jump describes a line that is no longer being answered for.
One product id, two registry lines: Base Set Venusaur as we read it
| Day | Total graded | PSA 10 |
|---|---|---|
| Aug 8 | 311 | 7 |
| Aug 12 | 311 | 7 |
| Aug 21 | 312 | 7 |
| Sep 8 | 314 | 8 |
| Sep 15 | 46,802 | 588 |
The rule we apply is mechanical: cut the series at every point that contradicts population physics, keep only the segment since the last cut, and if that segment holds a single reading, publish nothing. One reading on a fresh line is a value we watched replace another one, and we have no way to say which of the two belongs to the card. 3,476 lines are in exactly that state today.
That is not a hypothetical harm. Before the rule existed our screener printed Base Set Charizard at 105,288 and Blastoise at 51,522 beside Venusaur at 314 and Mewtwo at 363, four cards from one set whose registries cannot credibly differ by three hundred times. Every one of those four numbers looked like data.
Why two cards can report one population
The second failure is not staleness, it is attribution. The feed sometimes answers two different product ids with the same registry line, and when it does, both cards report the same population on the same day, down to the integer.
Identical readings served against two product ids
| Card | Reading | Also served to |
|---|---|---|
| Mew ex, 151 | 27,400 graded / 13,948 PSA 10 | product 519481 |
| _____'s Pikachu, Celebrations | 42,958 / 26,645 | product 161750 |
| Charizard SWSH066 | 9,003 / 4,985 | product 224366 |
| Base Set 2 Pikachu | 11,792 / 1,435 | product 161751 |
The most striking version is the Arceus set. AR1, AR2, AR4 and AR9 each read around 600 graded copies, and then all four returned 36,048 on the same day. Four distinct cards cannot share one lifetime submission count. One of them might own that line, or none of them might, and our rule can detect the collision without resolving it, so all four are withheld. 774 lines are in that state, covering 701 catalogue cards.
What withholding costs
4,689 of our 21,974 priced report rows publish no population figure, and those cards average $17.22 raw. They are not the expensive end of the catalogue, which is the one piece of luck in this: the cards where a population figure moves the most money are mostly cards whose line has been stable long enough to publish. But 4,689 is not a rounding error, and anyone using our grading pages should know that a card can be silent here while a competitor's page shows a confident number for the same card.
That competitor number is not better information. It is the same feed with the same collisions and the same re-pointings, printed without the checks. The gap between 19,950 and 24,200 is the part of the feed that does not survive being asked whether it is internally consistent.
Why the same rule protects the studies
Every population study we publish runs on the publishable set, not the raw one, which is why their conclusions are narrower and more durable than they would otherwise be. The gem rate curve and hard to gem both exclude these lines, and our study of thin registries would have been meaningless with 3,476 identity changes left in, because a line that just jumped from 314 to 46,802 arrives in the sample as a deep registry with a cheap card attached, which is the exact shape that study was measuring.
One more rule sits alongside these two and is worth naming because the Arceus example shows it plainly: a missing PSA 10 breakdown is not a PSA 10 count of zero. Those four Arceus readings arrived as 36,048 total with no gem figure, and an earlier version of this pipeline recorded them as a 0.0000 gem rate, which then scored as the most attractive grading candidate in the catalogue. Absent and zero are different values, and 5,874 feature rows once carried the wrong one.
Read from moonstone production on September 17, 2026, over the 400-day population snapshot window, which holds 24,200 distinct registry keys with a positive total. Publishable means the key survives both rules: the continuity rule (cut the series at any reading that falls by more than 1% or multiplies by 3 or more, then require at least two readings on the surviving segment) and the shared-line rule (readings of 500 or more graded copies matching another key within 2% on total and 0.02% on gem rate are withheld on both sides). Counts are 19,950 publishable, 3,476 withheld for identity change, 774 withheld as shared, and 0 withheld for any other reason. The priced-row figure counts the 21,974 rows of our report feature table carrying a positive raw price, of which 4,689 carry no population, averaging $17.22 raw. Example series are the stored daily readings for those product ids, quoted unmodified. We hold no information from PSA beyond these readings, so which card truly owns a contested line is not a question we can answer, only one we can detect.