Verification
Every class of fact on this reference is checked by a documented procedure before it is published, and each procedure records what it threw out. This page publishes those procedures and those rejection rates. The numbers below are computed from the run artifacts rather than written by hand, so if one is wrong the script is wrong.
An accuracy policy is easy to write and impossible to check. A rejection rate is neither. We publish ours because it is the part a reader can actually use to judge whether to trust the rest of the site.
| What was checked | Examined | Published | Rejected |
|---|---|---|---|
| Botanical identity | 256 | 108 | 148 (58%) |
| Research citations | 1251 | 496 | 755 (60%) |
| Safety data, first attempt (rejected) | 83 | 0 | 83 (100%) |
| Safety data, second attempt | 25 | 1 | 24 (96%) |
| Safety data, third attempt | 24 | 1 | 23 (96%) |
| Safety data, fourth attempt | 23 | 3 | 20 (87%) |
| Duplicate entries | 33 | 9 | 24 (73%) |
Why the rejections matter more than the acceptances
Two of these runs found that material already on the site was wrong. Every research bullet the encyclopedia carried was a paraphrase with no identifier, so no reader could check any of it; most of those were deleted rather than dressed up. Botanical identification rejected more candidates than it accepted, because a large number of classical drug names genuinely map to more than one species and picking a side would have manufactured a certainty the sources do not have.
The third run rejected itself. A panel of four reviewers produced safety records for the 83 highest-risk entries and passed 25 of them. An independent auditor then re-read a sample of those 25 and rejected every one, for defects including a dropped instruction to monitor liver enzymes, a source misread as saying the opposite of what it said, and a study's adverse findings reported as absent. Nothing from that run was published. It is listed here because a verification page that only reports its successes is not a verification page.
A blank field is a worse-looking page and a better reference. Where a page is missing something, it says so.
The standing rules
- Rejection is the default. Verifiers are instructed to refute, not to confirm, and an unresolved check fails.
- Verifiers are independent of identifiers. The agent that proposes a fact never approves it.
- Different lenses, not repeated ones. Where a fact can fail in several ways, each checker is given a different failure mode to hunt, because four identical reviewers find one class of error four times.
- Stated gaps beat silent ones. "No authoritative source establishes this" is a publishable answer. An empty section is not.
- The checkers get checked. A sample of accepted records is re-reviewed by an auditor who did not see the original verdicts.
Machine-readable
The full ledger, including per-record outcomes and reasons, is at /verification.json under CC BY 4.0.