Safety data, second attempt
Can the rejected records be repaired to a publishable standard? The 25 records the first panel accepted, re-checked against ten rules derived from the defects the meta-judge actually found in that run. Every repaired record was then audited individually rather than sampled, and 1 of 25 passed.
How it was checked
- Ten rules were written from the defects the first audit found, each one naming the real failure it came from: dropped monitoring instructions, claims attributed to studies that never tested them, adverse findings reported as absent, animal doses presented without species, spliced study arms, and developmental neurotoxicants graded as a caution.
- One agent per record re-fetched every cited source and either repaired the record or declared it unsalvageable.
- Every repaired record was then audited by an independent reviewer instructed to reject by default and to re-fetch the sources rather than trust the repair.
- The auditors also caught repairs that made records worse. On one, the repairer deleted a report of three deaths from liver failure as a fabrication; the auditor re-read the source, found the deaths were genuine, and rejected the repair for removing them.
What came out
| recordsExamined | 25 |
|---|---|
| published | 1 |
| heldBackForFurtherWork | 24 |
| unsalvageable | 0 |
| inconclusive | 0 |
Worked rejections
The rate says less than the reasoning does. These are real cases from this run.
abhraka-bhasma- RULE 6 (spliced study arms) + unsupported claim: the record's central hepatotoxicity finding states ALT 'rose to 132.5 IU/L (medium dose) and 127.8 IU/L (high dose) versus 93.3 IU/L in vehicle control'. Pandit et al. Table 7 reads: I-Control (Vehicle) 96.9 +/- 6.51; II-KVB-LD 93.3 +/- 7.99; III-KVB-MD 132.5 +/- 17.75**; IV-KVB-HD 127.8 +/- 5.96**. 93.3 is the LOW-DOSE arm, not the vehicle control. The record presents one treated arm's value as the control comparator in the single most extractabl | RULE 1 / RULE 4 (dose-extrapolation basis fabricated rather than dropped): doseLimits.text states
akika-pishti- Rule 2 (attributing to a study a statement it did not make): adverseEffects[1] presents as a verbatim quotation, inside quotation marks, "Ayurvedic lead poisoning produces greater hematopoietic toxicity than paint-removal poisoning." That sentence does not exist in Kales et al. I pulled the raw article text from PMC2538618: the Conclusions sentence reads "Ayurvedic poisoning produces greater hematopoietic toxicity than paint-removal poisoning." (no word 'lead'), and the only sentence containing | Rule 3 (a cited study's adverse findings must be reported): the heavyMetals section reports Bhall
ashwagandha- Rule 1 (selective quotation / dropped source directive): PMC13436050 states verbatim 'The reports necessitate monitoring liver function while taking Ashwagandha. Additional clinical trials are needed to determine the causality and safe dose levels.' The record carries no liver-function monitoring recommendation anywhere. It goes further and inserts, in the hepatic-impairment contraindication, 'LiverTox does not state a specific monitoring schedule and none should be inferred' - language that dis | Rule 3 (cited study reports adverse findings that the record does not report): LiverTox NBK548536
bhanga- Rule 1 - Selective quotation: LactMed motor-development sentence truncated mid-quote, dropping 'during the first month of lactation' and the 'number of reported joints per week' dose metric while retaining quotation marks | Rule 1 - Selective quotation: LactMed's 'in nonpregnant, nonlactating women' population qualifier deleted from the luteinizing hormone / prolactin finding presented in the lactation field
eranda-moola- Rule 1 (selective quotation is misrepresentation): the reproductive section is built from the single source that found nothing while root-specific reproductive toxicology is omitted, and the EMA source's own caveat that 'adequate tests on genotoxicity and carcinogenicity have not been performed' is not carried in the record body. | Rule 3 (documented adverse findings must be reported): root-specific adverse findings available on a permitted host are absent - 100% reduction of fertility index and suppressed testosterone in male mice from a root-derived protein (PMID 21288636), reversible antife
eranda-patra- Rule 1 — selective quotation: Mtendwa 2026 Table 2 assigns alpha-thujone to "Leaves, seeds". The record reports it as "tabulates alpha-thujone under leaves", dropping the seed co-attribution. In a record whose entire discipline is part-specificity (it correctly refuses to transfer seed-oil data to the leaf everywhere else), narrowing a two-part table row to the leaf alone is the same error in the opposite direction. | Rule 9 — documented substance-specific hazard with no contraindication of its own: contraindications is empty and interactions is empty, while the record itself documents, for th
Caveat
A pass rate of 1 in 25 is the honest result of applying a standard this strict to a literature that is mostly silent at the level of the individual preparation. The 24 held records are not discarded; they are unfinished.