Safety data, first attempt (rejected)
Can a four-lens judge panel produce publishable safety records? 83 highest-risk pages: 66 metallic/mineral preparations and 17 toxic or scheduled plants. The panel accepted 25 records. An independent auditor then re-reviewed 3 of those acceptances and rejected every one. None were published.
How it was checked
- Records were researched against an allowlist of regulatory and institutional sources, then judged by four independent lenses: source fidelity, clinical accuracy, regulatory language, and omission.
- A record shipped only if source fidelity passed it and at least two of the other three lenses agreed.
- A meta-judge then re-audited a sample of the ACCEPTED records, since a systematically lenient panel is invisible from inside its own verdicts.
- The meta-judge agreed with the panel on 0 of 3 sampled records.
- On that result the entire run was withheld. A defect rate that high in a sample means the accepted set as a whole cannot be trusted, and safety pages are the wrong place to publish and correct later.
What came out
| pagesExamined | 83 |
|---|---|
| Passed the judge panel | 25 |
| panelRejected | 49 |
| inconclusive | 9 |
| Accepted records re-audited | 3 |
| metaAgreedWithPanel | 0 |
| published | 0 |
Worked rejections
The rate says less than the reasoning does. These are real cases from this run.
abhraka-bhasma- MISLEADING DOSE FRAMING (most serious). doseLimits says "No human dose limit ... was located" and the note calls the figures "rat doses ... not a human dosing recommendation." Pandit et al. actually state the low dose (22.5 mg/kg) is the "highest clinical dose extrapolated to rats," that "one day in rat is supposed to be equal to 30 days in human" (so 28 days ≈ 2.5 human years), and that "In chronic asthma, KVB treatment is generally given for three months." The record strips all of this. As written, a clinician or an AI reading it concludes the hepatotoxic doses (45 and 100 mg/kg) sit far above human exposure. They do not: they are roughly 2x and 4.4x the extrapolated maximum clinical dose.
mrigashringa-bhasma- UNSUPPORTED CLAIM (verified against source). The record states: 'NCCIH notes that in a 2015 survey about one in four Ayurvedic supplements tested had high levels of lead and almost half had high levels of mercury.' The NCCIH page says two separate things: (a) 'A 2015 published survey of people who use Ayurvedic preparations showed that 40 percent had elevated blood levels of lead and some had elevated blood levels of mercury' — a survey of PEOPLE, not products; and (b) 'About one in four of the supplements tested had high levels of lead and almost half of them had high levels of mercury' — an undated statement about tested supplements, not attributed to the 2015 survey. The record has fused
varatika-bhasma- MISREPRESENTS A CITED SOURCE (v3, PMC7685255, verified by direct fetch). The record states 'No adverse effects reported' and quotes only the two exculpatory conclusion sentences from the Bacnil study. The study actually reports dose-dependent target-organ findings: at TED x10 (1962 mg/kg) mild-to-moderate liver fatty degeneration, cell infiltration and sinusoidal inflammation, mild renal fatty change, cardiac fatty change and gastric epithelial erosion; at TED x5 (981 mg/kg) liver fatty change and gastric epithelial erosion. Most reversed during recovery. Omitting the only positive toxicology signal in the only repeated-dose study that contains this ingredient is selective quotation, not ins
Caveat
The meta-judge re-audited 3 of the 25 panel-accepted records and rejected all 3, each for a specific and serious defect: a dropped monitoring instruction, a misread source, and an adverse-effect finding reported as absent. A 0 of 3 agreement rate means the panel's acceptances cannot be trusted as a set, so none were published.