40%
Safety
Reported tolerance, warnings, risks, and adverse-event context
PI / MethodologyPanacea Index / Methodology
Transparent by design
From paper to signal. Every step visible.
Panacea Index structures study-level findings into two inspectable views: directional consensus and weighted research aggregates.
- Method version
- 01
- Primary unit
- Study claim
- Direction classes
- 3
- Output
- Signal + confidence
The research path
One question. Five accountable steps.
Each stage narrows the distance between a broad supplement claim and the evidence that can actually support it.
- 01
Define the question
Start with an intervention and a specific outcome. Population, dosage, and time horizon are retained when the source reports them.
Output — A claim-sized research question
- 02
Structure the literature
Studies are connected to normalized supplements and outcomes, then classified as human, animal, molecular, clinical, or other evidence.
Output — A comparable evidence set
- 03
Extract study claims
Reported findings are represented as increase, decrease, or no clear effect, with the studied population and effect context kept alongside them.
Output — Traceable directional claims
- 04
Appraise and synthesize
For supplement aggregates, design, quality, citations, and recency inform study influence. For intervention–outcome pairs, directional claims determine agreement and conflict.
Output — Signal, strength, and disagreement
- 05
Expose the limits
The result is presented with its study count, confidence language, and source context. Sparse or conflicting evidence is labeled as such.
Output — An answer that can be audited
The unit of analysis
Consensus belongs to a pair—not a product.
A supplement can improve one outcome, have no effect on another, and carry different evidence across populations. The primary consensus view therefore asks about one intervention–outcome pair.
Intervention
Supplement
Endpoint
Health outcome
Population · dosage · duration remain attached when available
Pair consensus
Direction first. Strength second.
Each claim is classified as increase, decrease, or no effect. The leading class determines direction; its share of all claims determines agreement. Claim count sets the evidence floor.
Agreement
max(I, D, N)
I + D + N
I = increase claims · D = decrease claims · N = neutral claims. Agreement is rounded to a whole percentage for display.
Strong consensus
≥80% agreement
At least 20 claims
Moderate consensus
≥60% agreement
At least 10 claims
Emerging evidence
A leading direction
Below stronger thresholds
Conflicting evidence
<50% agreement
No clear majority
Insufficient evidence
Fewer than 3 claims
Too little to classify
Study influence
Papers contribute differently.
The aggregate model gives more influence to higher-quality, more direct, better-established, and more recent evidence.
Final study weight
quality × citations × study type × recency
quality = 1 + max(0, score − 50) / 100
citations = 1 + log(count + 1) / 10
recency = exp(−years_old / 15)
weight = all four factors multiplied
Clinical or human
2.0×
Most direct evidence for human decisions
Animal
1.5×
Mechanistic and preclinical context
Molecular
1.2×
Biological plausibility and mechanism
Other
1.0×
Baseline contribution
Supplement aggregate
Three dimensions, kept distinct.
Safety, efficacy, and study quality are normalized to a 0–1 scale, combined at the paper level, then aggregated using each study’s calculated weight.
40%
Efficacy
Reported effectiveness for the studied intervention and outcome
20%
Quality
Design strength, sample, duration, and study-level appraisal
Paper score
(safety × 0.40) + (efficacy × 0.40) + (quality × 0.20)
Weighted aggregate
Σ(metric × study weight) / Σ(study weight)
Aggregate confidence
Confidence requires both a minimum number of studies and enough total weighted evidence. Missing either condition lowers the label.
High
≥20 studies
≥30 total weight
Medium
≥10 studies
≥15 total weight
Low
≥5 studies
≥5 total weight
Very low
<5 studies
or <5 total weight
Consistency = max(0, 1 − standard deviation / mean)
AI-assisted, source-grounded
Automation helps read at scale. It does not confer truth.
AI assists with structuring study text, identifying claims, and appraising fields such as safety, efficacy, and quality.
Those outputs inherit the limits of the source and the model. Important findings should be checked against the cited paper.
Known limitations
What the model cannot settle.
Methodology is not a substitute for judgment. These constraints should travel with every result.
A score is not a clinical probability
A 0.8 score does not mean an intervention has an 80% chance of working. Scores organize evidence; they do not predict an individual response.
Citation count is an imperfect signal
Older, popular, or controversial papers may accumulate citations for reasons unrelated to methodological quality.
Recency is not the same as quality
Newer evidence receives a recency adjustment, but a recent weak study does not automatically become strong evidence.
Direction compresses effect size
Increase, decrease, and no-effect labels make studies comparable, but they do not by themselves communicate magnitude or clinical importance.
Missing data changes the picture
Incomplete abstracts, absent population details, and unavailable full text can reduce the specificity of extracted claims and scores.
Automation can be wrong
AI-assisted extraction and appraisal may misread a paper. Source-level details should be checked before making consequential decisions.
Read critically