The claim this platform makes is that its numbers can be audited. That is a comfortable thing to say and an uncomfortable thing to test, so we tested it: a checker that reads every article, extracts every figure and every link, and recomputes them against the live database rather than against what the article says. We ran it on all 133 published pieces. It returned 133 findings. This is what they were.
The dangerous error is not a typo
A mistyped number is caught by proofreading. The error that survives proofreading is the one where every figure is computed correctly from the wrong starting point. Our worst case: a set's recorded history usually begins on the day we added it, not the day it was released, so measuring "since launch" from the first reading measures our onboarding. On Destined Rivals that reads as a 108.3% gain. Measured from the first reading after the set was actually on sale, it is a 9.6% loss. Two drafts arguing the first version were written, checked, and pulled before publication.
One published article did not get that treatment in time. It compared nine sets over their "first ninety days", and for six of them the window was one shared stretch of early 2025 on sets that were three to thirty-nine months old. It has been withdrawn, and the ten citations to it in eight other articles have been rewritten to say what was actually measured.
Four rules, each one earned
The rules, and the failure that produced each
| Rule | What it catches | Found |
|---|---|---|
| Release anchor | A launch or release claim about a set whose series begins after the set came out | 24, in 10 articles |
| Window length | A window longer than the tape behind it, which a clamping helper silently returns as something else | 6 |
| Link resolution | Card, set and article links that do not resolve | 103 |
| Superlatives | Reported, never judged: only a recompute settles "the only" or "every" | listed per article |
The window rule is worth explaining because it is the subtlest. A ninety-day change was published for a set with seventy-six days of history. The helper that fetches "the reading ninety days ago" does not fail when the series is shorter, it returns the earliest reading it has, so the article printed the set's entire lifetime change under a ninety-day label. Nothing about the output looks wrong.
A hundred doors that open onto nothing
The largest single finding is not about numbers at all. A hundred internal links across the corpus point at set pages that return 404, across thirty-six distinct slugs, and they cluster on promo bins and older lines. Internal links are what tells a reader, and a crawler, which of tens of thousands of near-identical catalog pages are worth visiting. A hundred of ours point into a void.
What this does not fix
A checker can verify that a figure reproduces and that a window is real. It cannot tell that a sentence like "the deepest comp pool of any card we rate" is false, because that is a claim about a comparison nobody recomputed. Those are caught only by an adversarial pass over the prose, and in our experience that is where the errors actually live: in three articles checked this month, roughly 180 computed figures reproduced to the cent and most of the errors were in the sentences written about them rather than in the arithmetic.
The checker runs against production data, not against a fixture, and is wired into the deploy so a new finding blocks a release. It reads every block of every article, extracts card, set and article links and every percentage and dollar figure, and pairs them with the set or card they sit beside. Findings are recorded in a baseline file with their exact text, so a finding that is fixed disappears and a finding that is new fails the build; the baseline is a debt register and is expected only to shrink. Article audits are recorded against a content hash, so editing an audited article invalidates its audit and blocks the deploy until it is rechecked. Counts in this piece are as of September 21, 2026. Nothing here is investment advice.