How scoring works
Every Avelor verdict is a 0–100 composite of five axes. The same five inputs power every score, every time. No paid placement, no editorial reranking. The inputs are deterministic given the product label and the user's profile.
Full source-list at /our-data. Full benchmark + grader rubric at /methodology.
Ingredient evidence
Per-ingredient evidence grade (A / B / C) drawn from RCT count, meta-analysis presence, and NIH ODS classification. Ingredients with strong-evidence claims at the trial-effective dose contribute positively; ingredients with debunked or limited evidence reduce the score.
Source: 772+ researched ingredients · NIH ODS · PubMed
Dosage adequacy
Each active ingredient is compared against the minimum effective dose from clinical trials. Underdosing is the most common reason a scan loses points. Hidden 'proprietary blends' are flagged when the per-ingredient amount cannot be verified.
Source: Per-ingredient clinical-trial dose ranges from the research corpus
Drug-supplement interactions
When the user's profile lists medications, every ingredient is checked against the verified interaction database. Severe interactions (e.g. 5-HTP + SSRIs, St. John's wort + birth control) cap the score regardless of other axes.
Source: 757 supplement and medication pairs reviewed · 317 unique PMIDs
Third-party test signals
Pulls from independent lab-test data, including products carrying the USP Dietary Supplement Verification Program or NSF Certified marks. Products that have failed third-party content or contamination tests are penalized. Products without any third-party data carry a small uncertainty discount.
Source: 373 failed products in surfaced corpus · independent lab testing
Recall + adverse-event flags
Cross-references the FDA recall and CAERS adverse-event databases. Products in active or recent recall windows are flagged. Brands with disproportionate adverse-event signals carry a brand-level penalty.
Source: 2,354 FDA recalls · 2,061 tainted-product records · 772 CAERS events
Severity scale
Documented serious-adverse-event or contraindication risk. Combinations like 5-HTP + SSRI (serotonin syndrome) or red yeast rice + statin (rhabdomyolysis). Verdict caps + a clear safety-first surface.
Clinically significant interaction or absorption issue with established literature. Examples: levothyroxine + calcium (4-hour separation), warfarin + high-dose fish oil (bleeding risk).
Mechanism-plausible interaction with limited or mixed clinical data. Worth flagging; not a stop sign. Surfaces a discuss-with-prescriber prompt without overstating risk.
Theoretical or rarely-clinical-grade interaction. Logged but not surfaced as a primary warning unless the user explicitly asks.
Evidence tiers
Every claim backed by a Avelor interaction or evidence grade is tagged with the strongest tier of supporting evidence we can verify. Higher-tier evidence weighs more heavily in the final score.
- Systematic review / meta-analysis: pooled evidence across multiple RCTs. Highest weight.
- Randomized controlled trial (RCT): single-trial evidence with proper design.
- Observational / cohort: population-level association without randomization. Useful for prevalence and signal detection.
- Case report: single or small-series clinical observation. Sufficient to flag but not to rule.
- Mechanism only: biochemical or pharmacokinetic plausibility without published human data. Lowest weight; usually surfaces as a low-severity note rather than a recommendation.
- Label / regulatory: claims sourced directly from FDA / DailyMed / NIH ODS labelling. Treated as authoritative for the specific claim made on the label.
Confidence + insufficient evidence
Each scan returns a numeric confidence score derived from how many of the five axes had high-tier supporting evidence. When the evidence is genuinely contested, mixed, or missing, Avelorsays so, explicitly, instead of generating a confident-sounding answer.
- High confidence: at least 3 of 5 axes have RCT or systematic-review-grade evidence, and there is no contradicting source. Verdict ships with a clear numeric score.
- Moderate confidence: evidence is mixed across axes; verdict ships with the score plus an explicit "based on partial data" caveat.
- Insufficient evidence: the system does not invent a score. The user sees an honest "we don't know yet, here's what is and isn't in the literature" surface and is pointed at the upstream sources (NIH ODS, NCCIH, DailyMed, PubMed) for self-review. We never guess.
This rule is consistent with the safety-critical-omission hard-fail in the public benchmark. See /benchmark for the rubric.
What we deliberately don't do
- No paid placement. Brand pages do not buy a higher score.
- No editorial reranking. The score is deterministic given the label + profile inputs.
- No personalized score boost from affiliates. Affiliate revenue, when present, is disclosed at the product level and does not feed back into scoring.
- No claim of medical advice. The score is an information aid; decisions about supplement use belong with the user and their prescriber.
Snapshot date: 2026-09-25. Weights re-tuned at most quarterly with version notes published at /methodology.