August 10, 2026
We Fixed The Definitions And Missed Our Own Explainer
Our badge explainer said Source Diversity asks whether reviews come from multiple platforms and Claim Verification asks whether a brand's claims check out. Neither is what the audit scores.
Earlier today we published a correction saying our methodology page described five measurements we do not perform. We fixed that page, the trust page and the machine files in one release. Then we left our own badge explainer teaching the definitions we had just retired.
The explainer is where a reader goes to find out what the green shield means. It said Source Diversity asks whether reviews come from more than one platform. It said Source Quality means a verified Amazon purchase outranks an anonymous forum, that Claim Verification asks whether a brand’s claims check out, and that Temporal Consistency asks whether quality held up over the years.
Not one of those is a question the audit asks. Source Diversity scores where the reviews live and whether that place can be trusted to hold real ones. Claim Verification scores customer photographs. Temporal Consistency scores whether anyone discusses the product somewhere we have no influence at all. The page now carries all five question sets with the points each one awards, so the badge can be checked against the same list we score it from.
The wider measurement is the part worth publishing. On 10 August 2026 we read every sentence in our live library that explains a factor score. 409 of them, spread across 164 records, justify a score using a measurement that factor does not perform. They read the new names literally, which is the exact trap our own scoring spec was rewritten to close.
We are not rewriting those sentences, and the reason matters more than the count. A justification is the reason a number exists. Swapping the reason while keeping the number would invent a derivation nobody ever ran, and inventing one is the single thing this system is built never to do. Every one of those records is queued for a re-score against the real questions, and each correction publishes when its re-check lands, not before.
What we are keeping from this: a correction that only fixes the pages it was written about is half a correction. The first one counted four published surfaces. It did not count the records quietly repeating the same thing, one product at a time, and our first count of those was low as well.