Where every number comes from
Every fact in Litmus traces to a named public source, and this page is the supply chain. It is the same catalog we hold ourselves to internally, published, including the parts a marketing page would normally leave out.
The sources
Product labels and their ingredient lines. The catalog: every product page starts as a DSLD label, loaded in bulk and upserted deterministically.
What the evidence says per nutrient. The Health Professional fact sheets; our evidence summaries are extracted from these, with the source URL kept on every row.
Research paper metadata and citations. We store the citation (PMID, title, journal, authors, link) and never the abstract or full text: those belong to the publishers.
Citation-impact ranking for papers. Field- and age-normalized impact, so a flashy pilot study cannot outrank a meta-analysis.
Product recalls. Recall facts matched by exact firm name only: fuzzy matching a recall to the wrong brand is a bell you cannot un-ring.
The outcome vocabulary. Health outcomes are named with real Medical Subject Headings descriptors, not labels we made up.
The counts
products · from the DSLD bulk load
ingredients · name-normalized, synonym-merged
label lines · one per ingredient per label
papers linked · PubMed citations, iCite-ranked
evidence rows · all but one carrying the exact source URL
recall facts · openFDA, exact firm match
Counts as of the last catalog reconciliation, 2026-07-05. When the catalog changes, this page changes in the same release.
What involves AI, exactly
Honesty requires precision here, so: two things in the Litmus data involve a language model, and neither invents facts.
- Evidence summaries are AI-extracted, from cited sources. The plain-language summary of what the research says for an ingredient is extracted by a model from that ingredient’s official NIH ODS fact sheet: extraction, not free generation. Every one of these rows keeps the exact source URL and its review date, so each is auditable against the document it came from. A minority of rows are hand-written from named sources.
- Goal-to-outcome mappings are AI-picked, from a fixed vocabulary. Matching your goals (sleep, energy) to formal research outcomes was done by a model choosing among real MeSH descriptors: it could select, not invent. Where no confident match existed, we left it unmapped rather than guess. A few later mappings were added by hand, human-reviewed against the same vocabulary.
Everything else, the catalog, ingredients, brands, papers, and recall facts, is deterministic ingestion of the public sources above. No model writes product data.
What we do not know yet
- The catalog is a snapshot: products added to DSLD since our last bulk load are not in yet.
- Barcode coverage is partial (roughly half of products carry a UPC in the source data).
- Evidence quality grades exist internally but are not shown publicly until our grading is fully deterministic rather than model-assigned.
- Evidence coverage skews toward vitamins and minerals (the shape of NIH fact-sheet coverage); expanding it is active work.
- One evidence row (apigenin and sleep) currently cites no source URL: the paper it used to cite was off-topic, so we unlinked it and marked the row insufficient until a valid human study is cited. Honest gap, kept visible.
The commitments
- No affiliate links, no house brand, no sponsored rankings: nothing to sell you but the software.
- No invented facts: if we cannot source it, we do not show it.
- Recalls are reported as recorded by the FDA, matched conservatively.
- When the honest answer is “the evidence is thin,” that is the answer you get.