Method

Track record is calibrated against FDA's quarterly posting of potential signals, required by FDAAA 2007 section 921 — 570 postings over 23 quarters covering 343 distinct products. No product monitored here appears among them. Nor does any vaccine, blood product or gene therapy, which says something about the postings we hold rather than about FDA: this is the tracking system covering small molecules and monoclonal antibodies, not the one covering plasma-derived products. FDA does act on plasma products, through CBER channels this system does not ingest. Seven such actions on these classes are recorded in signal.regulator_action, from a boxed warning for intravascular haemolysis on Rho(D) immune globulin in 2010 to immunoglobulin lot withdrawals for hypersensitivity in March 2025. That register is hand-entered from web search and is the only thing on these screens no query can verify: 0 of its 7 rows have been checked against source by a person. So the finding that these statistics separate FDA's actions from its declines by only 1.296x is measured on a different population and carried across, which assumes the relationship holds for plasma-derived products. That assumption has not been tested here. The precise statement is the narrow one: no monitored product appears in the 570 postings ingested.

· calibration_set_contains_no_monitored_product

A reaction reported for every drug is invisible to a ratio, however grave it is. Renal dysfunction is a BOXED WARNING on immunoglobulin and ranks 209 here, with a PRR of 0.675 — below one, meaning immunoglobulin is under-represented among renal failure reports relative to the rest of the corpus. It has 254 cases, but they are 0.28% of the 91,520 renal failure reports in FAERS. Aseptic meningitis ranks first because this drug accounts for 26.7% of every aseptic meningitis report there is. Both facts are true and only one of them is a signal a ratio can find. This is the reason CIOMS has five steps rather than one, and the reason the watchlist tracks concepts every quarter regardless of where they rank.

· disproportionality_finds_distinctive_not_serious

Measured against 289 graded FDA adjudications: 82 of 204 label changes had no usable FAERS evidence before FDA acted, and 51 of those had literally zero cases in the public data at the time. FDA sees manufacturer submissions this corpus does not contain. No arrangement of public data closes that gap, and a tool claiming otherwise would be claiming to see what is not there.

· public_faers_ceilings_around_sixty_percent

FAERS records reports, not patients treated. Nothing here says how many people took a drug, so a rate per exposed patient cannot be derived at any point, by any method, from this data. Every figure in this system is a proportion of REPORTS. A product whose reporting rises may be more used, more scrutinised, newer, or genuinely more harmful, and disproportionality alone cannot separate those.

· spontaneous_reports_have_no_denominator

Twenty-two fields are recomputed by signal.verify_method_evidence() on every consistency sweep — including the register's verified count, so confirming a source makes the prose above false and turns the sweep red until it is corrected.

Every number quoted in the limitations above is stored as evidence and recomputed against the live database. A figure in prose that goes stale breaks a check here rather than quietly misleading — which is how the migration counts, the operating curve and the register status have all been caught.· signal.verify_method_evidence()