The Magellan Longevity Open Evidence Dataset

What this is

Magellan Longevity grades every product it lists on the strength of the human evidence behind it, and links each grade to the primary literature it rests on. This dataset is that entire judgement, exported in full and given away: 130 products with their grade, the reason for the grade, the caveat that qualifies it, the dose actually used in human trials, the mechanistic pathway, and every PubMed identifier behind it — plus a second table of 2692 study-level records covering 2646 distinct PubMed identifiers and 2346 DOIs.

It exists because evidence grading in consumer health is almost always a black box. Sites publish a badge and never publish the reasoning, the citation list, or the cases where the evidence went the other way. Everything here is machine-readable and auditable, including the 1 product graded “Tested — did not work” and the 144 study records whose own conclusion reports no difference. Null results are part of the dataset by design — a catalogue that only ever surfaces flattering studies is marketing in a lab coat.

The grades in this file are not a re-derivation for publication. They are produced by executing the exact grading code that renders the live product pages, so the dataset cannot drift from what a reader sees on the site.

Download

Every file is UTF-8, licensed CC BY 4.0, and free. CSVs follow RFC 4180. The two tables share an identical schema across all three formats — the JSON and JSONL are not reduced versions of the CSV.

Supporting files: data-dictionary.json · data-dictionary.md · dictionary as a web page · datapackage.json (Frictionless) · dataset-metadata.json (schema.org) · CITATION.cff · citation.bib · citation.ris · LICENSE.txt · checksums.txt · README.md

# verify what you downloaded
sha256sum -c checksums.txt

# 10 strongest-evidence products, straight from the CSV
python3 -c "import csv;[print(r['evidence_grade'],r['product_name'])\
 for r in csv.DictReader(open('products.csv')) if r['evidence_tier_key']=='strong']"

What is in each table

Table 1 — products (130 rows × 32 fields)

One row per catalogue product. Beyond identity and category it carries evidence_grade, grade_rationale, grade_caveat, studied_dose and dietary_dose, pathway_name, citation_count, the semicolon-separated pmids, and study_keys to join into table 2. A verbatim key_finding_quote records the single sentence used to justify the pathway assignment — quoted, never paraphrased into a stronger claim.

Table 2 — studies (2692 rows × 32 fields)

One row per study record: pmid, doi, journal, year, title, the compound it supports, the outcome in the source's own words, and reported_direction — a coarse rule-based reading of which way the source itself says the result went. 314 rows also carry a full structured write-up: research question, hypothesis, methods, results, conclusion and stated limitations. Filter is_primary_record_for_pmid to reduce the table to 2646 unique papers.

Join: products.study_keys → split on ; → studies.study_key. The full field-by-field dictionary is at /data/data-dictionary.html.

How it was built

A single dependency-free Node script (tools/build_dataset.js) reads five files from the live site and writes every artefact on this page. Nothing is hand-edited between generation and publication, and no number on this page is typed by hand — each one is computed at build time.

Empty means not recorded in the source. No value is imputed, rounded up, or filled in from memory: where the catalogue has no DOI, the DOI cell is blank; where no human dose range has been established, studied_dose is blank rather than guessed.

The grading rubric, in full

Ingestibles and interventions are graded on the strength of human evidence for the active compound. Diagnostics and wearables are graded on how well the thing they measure predicts health — not on a treatment effect. These are the tier definitions exactly as published on the site, with the number of products that currently sit in each.

For well-characterized compounds, a board-certified physician assigns the grade from the human and mechanistic literature, and we publish a typical dose only where a verified standard range exists. For the long tail of ingredients, the grade is derived conservatively from the peer-reviewed studies we link on the page, and is capped below “Strong” until it has been individually reviewed. Effect statements shown on product pages are quoted verbatim from a cited study — we never paraphrase findings into stronger claims than the authors made.

We show null and negative evidence too. Where human trials are inconsistent or have failed to show benefit, we say so on the product page. A marketplace that only ever surfaces flattering studies is marketing in a lab coat.

Editorial firewall. Inclusion is never paid. No brand can buy a listing, a grade, or placement. We are not sponsored by and do not sell our own supplement line — so we are never stuck defending a bottle we make.

Each grade also carries how it was reached: curated_compound (104), curated_product (26). Open the full methodology in the Magellan store →

What the dataset says about itself

These are the aggregate numbers a reader or a journalist would want before citing anything. All are computed from the released files.

Distribution of evidence grades

Across the catalogue the median product carries 21 peer-reviewed citations; the total number of product-to-study links is 3421.

Direction of effect reported by the sources

Read this honestly. reported_direction is a deterministic keyword cascade over the source's own summary sentence, not an appraisal and not a language model. unclear (50.7% of rows) means no cue matched — usually a descriptive or mechanistic conclusion — and must not be read as a null result. Every classified row carries direction_evidence, the exact phrase that triggered it, so you can audit or discard any row. The rule set is documented in full in the data dictionary.

Publication years of the evidence base

Range 1971–2026, median 2019, across 2692 rows carrying a year.

Most-cited journals

1153 distinct journals appear in the dataset.

Mechanistic pathways covered

24 distinct pathways in total — the pathway taxonomy is rendered visually at /pathway-universe.html.

Preview

Fifteen rows from products.csv, strongest evidence first. The full table of all 130 rows is browsable at /data/evidence-grades.html.

Limitations you should know before citing this

• It is an evidence map, not a systematic review. There is no protocol registration, no exhaustive search strategy, no risk-of-bias instrument and no meta-analysis. It documents which literature this catalogue relies on and how strongly it was graded.
• reported_direction is derived, not curated. It is a published rule cascade over the source's own words. 1365 rows are unclear because no rule fired; that is not evidence of absence.
• 288 study rows are not yet attached to a compound monograph and therefore have an empty primary_compound. They are retained rather than dropped so the table reconciles exactly against the source catalogue.
• 46 rows are duplicate PubMed identifiers held under more than one internal key. Filter is_primary_record_for_pmid to collapse them.
• 346 rows have no DOI because the source catalogue records none. They are left blank rather than reconstructed.
• Prices are a snapshot from 2026-08-11 and are included for context only. Live prices change daily.
• Grades are product-level and evidence-level, never a recommendation. A “Strong evidence” tier says a specific measurable outcome is well supported for the compound — it does not say the compound extends life, and it says nothing about whether a given bottle contains what the label claims.
• 284 press and institutional news items in the source catalogue are excluded from table 2 on purpose: they are not primary peer-reviewed literature.

Licence and how to cite

Released under the Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt it for any purpose, including commercially. You must give credit, link to the licence, and indicate whether you made changes. In practice that means: if you use these grades or this citation graph, name the source and link back.

APA

Radu, G. (2026). Magellan Longevity Open Evidence Dataset (Version 1.0.0) [Data set]. Magellan Longevity. https://magellanlongevity.com/data/

BibTeX

@dataset{magellan_longevity_evidence_2026, author = {Radu, Gabriel}, title = {Magellan Longevity Open Evidence Dataset}, year = {2026}, version = {1.0.0}, publisher = {Magellan Longevity}, url = {https://magellanlongevity.com/data/}, license = {CC-BY-4.0}, note = {130 graded longevity products and 2692 linked study records} }

Plain attribution line

Magellan Longevity Open Evidence Dataset v1.0.0 (Gabriel Radu, DO, 2026), https://magellanlongevity.com/data/ — CC BY 4.0

Machine-readable citation metadata: CITATION.cff, BibTeX, RIS. Third-party article titles, journal names and short quoted findings remain subject to their publishers' terms; no full texts or complete abstracts are redistributed.

Versioning

v1.0.0 — 2026-08-11. First public release: 130 product rows, 2692 study rows, 2646 distinct PubMed identifiers. Schema frozen at 32 product fields and 32 study fields.

Releases are semantic: a patch changes values only, a minor release adds fields or rows, a major release changes or removes a field. Every release ships its own checksums.txt so a downstream user can prove which cut they analysed.

Related on Magellan