Standing page
How this publication counts
Every investigation here states its own method. This page states the rules all of them follow, so the per-piece methods box can stay short.
Where the numbers come from
Two sources, and no others. The first is this portfolio's own corpus: 77 compound records, 77 generated evidence profiles and 77 literature retrieval files, all of them JSON in the repository that builds these sites. The second is the US Federal Register's public API, queried live.
Nothing is drawn from a secondary summary, a vendor page, a press release or another publication's reporting. Where a figure could not be computed from one of those two sources, it does not appear.
Five rules, applied to every piece
One. Every figure is generated. No number is typed into prose. A single script, scripts/sentinel-analysis.mjs, computes every figure on this site into one file, and the templates read from it. A number that cannot be regenerated cannot be published.
Two. Every piece carries a methods box. What was counted, from which files, on what date, with what query, and where the count stops. It is a fixed shape, it sits in the body of the piece rather than in a footnote, and a piece that cannot fill it is not published.
Three. Every piece carries a limits section. Stated as what the analysis cannot show, not as a hedge. The limit that applies to all of them: no full texts were read, so nothing here is a claim about what any study concluded.
Four. Every piece links a dataset. Four tables, 833 rows in total, each with a page defining every column. If a reader cannot check the number, publishing it is an assertion rather than reporting.
Five. Where the subject is this portfolio, the piece says so. Three of the four investigations measure the corpus this portfolio itself publishes from. Each says that in its standfirst, in its methods box and in its dateline. Reporting on your own dataset as though it were someone else's is the one failure this publication could not recover from.
What is never claimed
No full text was read for any piece. Not one sentence on this site describes what a study found, what a rule requires or what the law is.
No derived label is presented as an assessment. Where this publication reports a tier, a category or a flag, it reports the test that produced it in the same table.
No compound is recommended, ranked, compared or appraised. This publication does not publish a page about a compound, and the one case where it comes close, a table of flagged titles, is labelled as the output of a screen on every row.
No comparison is made with other reference sites. None of them publishes the query behind its citation list, so there is nothing to compare against, and an unmeasurable comparison is worse than none.
Vocabulary
| Term | What it means here |
|---|---|
| Retrieved hits | The number of records a literature query returned on the date it ran, as the index itself reported it. A hit is a record matching a phrase, not a study about a compound. |
| Indexed rows | The records this portfolio's corpus actually carries for a compound, after the retrieval cap. Always twelve or fewer. |
| Coverage | Indexed rows divided by retrieved hits, for one compound, for one query, on one date. Not a measure of quality or of completeness in any absolute sense. |
| Evidence tier | A derived label produced by an ordered ladder of tests over counts of record types. Marked INFERRED wherever it appears, always reported with the basis sentence that produced it, and never an assessment of clinical efficacy, safety, study quality or consensus. |
| Study model | Whether a record describes work in a dish, in an animal or in people. Derived only from the subject headings and publication types the indexer assigned. Where those do not settle it, the record is recorded as unsettled rather than guessed. |
| Discordance candidate | A record whose title matched a negation marker. A candidate is a title, not a finding and not a judgement on a study. |
| Document, in the register | One item published in the Federal Register, identified by its document number. A count of documents is not a measure of regulatory burden, force or enforcement. |
Rounding, dates and corrections
Counts are integers and are never rounded. Percentages are printed to whatever precision keeps them meaningful: the smallest coverage figure in the corpus is under one hundredth of one per cent, so coverage is carried to four decimal places in the data and printed at the precision that does not imply false resolution.
Every count carries the date it was taken. Figures drawn from the Federal Register are a snapshot and will not reproduce exactly on a later date; the dataset carries the retrieval date so a later pull can be compared rather than confused.
A wrong number is corrected by fixing the script, regenerating the dataset and restating the figure in place, with a note saying what it was and when it changed. Numbers are never quietly edited, and a correction is never made by deleting the piece.
The counting rules on this page apply to everything this publication prints, including the pages describing itself. The editorial policy states what it will and will not publish, and the data section holds every table behind every figure.