Methodology
How Elivion
grades evidence
A rating you can't audit is just another opinion. This page documents exactly how every Elivion readout is produced — from the databases we query to the safeguards that keep the score honest.
Step 1 — Retrieval
Five public databases — plus our own index
Up to ~400 studies are collected per query, then de-duplicated at DOI and title level so the same trial can't be counted twice. In parallel, Elivion's own semantic index recalls every relevant study the engine has ever processed.
PubMed
The primary biomedical index of the U.S. National Library of Medicine. Queried via the official NCBI E-utilities interface.
OpenAlex
An open citation graph of 200M+ scholarly works. Lets us weigh how often a study is actually cited by later research.
ClinicalTrials.gov
The registry of interventional trials. Surfaces what is currently being tested — including trials that haven't been published yet.
Google Scholar
Catches grey literature and journals outside the core indices, so a narrow database selection can't skew the picture.
bioRxiv / medRxiv
Preprint servers for the newest findings. Always labelled as preprints — early signal, never treated as settled science.
Elivion Index
Our own, rapidly growing research memory. Every processed study is encoded by meaning and instantly retrievable for every future query — proprietary coverage that compounds with each analysis.
Step 2 — Classification
Not all studies count the same
Every study is typed into one of eleven evidence classes — detected from both title and abstract, so an RCT with a vague title is still recognized as an RCT. A meta-analysis of 42 trials outweighs a case report, and the grading reflects that.
Ranking then combines semantic relevance — does this study actually answer the question? — with citation impact and recency. The top 35 studies go into the grading step.
Step 3 — Grading
A calibrated 0–10 scale
Claude, Anthropic's frontier model family, weighs the ranked studies against an explicit calibration — the same yardstick for every treatment, from cryotherapy to supplements. Scores map to evidence levels: strong, moderate, emerging, sparse. And grading happens per clinical application, not per buzzword: one treatment can be strong for recovery and merely emerging for sleep — the readout shows where the literature agrees and where it thins out.
Step 4 — Safeguards
What keeps the score honest
AI accelerates the reading; it doesn't get the last word. Four layers sit between a model output and a published score.
01
Structured output, enforced
The AI cannot answer in prose. Every assessment must pass a strict schema — and every key finding must carry a concrete data point and its source. No citation, no finding.
02
Consistency checks
After grading, automated rules cross-examine the result: if the narrative describes multiple strong RCTs but the score doesn't reflect them — or vice versa — the score is corrected before anything is published.
03
Conservative anchors
Well-established interventions carry conservative minimum scores, so one thin retrieval run can never underrate settled science. Anchors only apply when enough studies were actually found.
04
Experts in the loop
Grading prompts are versioned like source code. Every change is tested in side-by-side comparisons and voted on by domain experts before it goes live — and every published run is stored with its exact prompt, model and study snapshot, fully reproducible.
Independence
Scores cannot be bought
Evidence scores and company verification are strictly separate systems. Verification levels attest that a company is real, documented and consistent — they never move an evidence score. A verified company with a moderately supported treatment stays at moderate. That separation is the product.
Limitations
What a readout is not
An Elivion readout summarizes published research — it is not medical advice, and it can't replace a clinician who knows your history. Science moves: assessments are dated, cached results expire, and preprints are always labelled as such. When the evidence is thin, the score says so instead of guessing.