Methodology
Version v1. Every score snapshot records the version that produced it, so historical charts stay interpretable when the methodology changes.
The division of labour
A language model reads each study and extracts what it says: design, sample, endpoint quality, risk of bias, replication status, and how the result sits against prior evidence. It never returns a score.
Deterministic code turns those extracted values into a score movement. The same inputs always produce the same number, which is what makes an evidence history auditable rather than a matter of opinion.
Humans handle ambiguity: disputed interpretations, low-confidence classifications, and every change large enough to matter.
What moves a score
| Study design | Weight |
|---|---|
| Meta analysis | 1.00 |
| Systematic review | 0.90 |
| Rct | 0.85 |
| Cohort | 0.50 |
| Case control | 0.35 |
| Observational | 0.30 |
| Animal | 0.15 |
| Preclinical | 0.12 |
| In vitro | 0.08 |
| Case report | 0.05 |
| Other | 0.05 |
Weight is then adjusted for sample size, follow-up duration, endpoint quality, risk of bias, replication, consistency with prior evidence, and effect magnitude. Preprints are discounted by half. No single event can move a score by more than 8 points.
Score labels
The headline score is a weighted composite: human evidence 60%, animal 20%, mechanistic 20%. Safety acts as a ceiling rather than a bonus, so an intervention with a poor safety profile cannot score highly on the strength of its mechanism.
What always requires a human
- Any proposed change of more than 5 points
- Landmark and major findings
- Anything that weakens the evidence base
- Safety signals, regulatory decisions, retractions, replication failures
- Any change to a high-profile intervention scoring above 80
- Any extraction below 97% model confidence
The Cancel Death Index
An experimental composite of five pillars, weighted as below. It is not a probability, a forecast, or a peer-reviewed instrument, and it should not be cited as one.
- Human trial
- 30%
- Regulatory
- 15%
- Translation
- 20%
- Biomarker
- 15%
- Healthspan
- 20%
Corrections
Every evidence page carries a correction link. Corrections go to a moderator queue rather than editing the record directly. An accepted correction earns its reporter reputation, triggers a re-analysis where one is warranted, and appears in the public change history for that entity.