Signal types and scoring.
Whel reads each drug-condition pair through three evidence arms, each pulling a different kind of source and held to its own scoring bar. Every arm is scored on the same five dimensions, but each dimension is interpreted on that arm’s terms, so a patient report is never judged on clinical-trial criteria. The arm scores are then discounted by how far the evidence was generated in women, and sorted into a confidence tier.
Three arms, scored separately and reported side by side.
A published trial, a mechanistic link, and a community report each carry a different kind of information and a different kind of error. Whel reads all three, scores each against its own bar, and reports them side by side. The strongest Direct reading anchors the headline, and the others corroborate beside it. They are never averaged into one number. Expand each arm below for its full scoring criteria.
Direct Research is the most literal arm: studies and trials explicitly designed to investigate the condition in question. It is intentionally sparse for most of these conditions, and the sparseness is itself data, reflecting how little targeted research exists. Sources are PubMed (via the NCBI Entrez API) and ClinicalTrials.gov (REST API v2, interventional studies). Every claim is pinned to a verbatim quote from the source and the offsets are verified independently of the model.
Scoring criteria · five dimensions, 0–2 each
The same five slots are scored for every arm, but each slot is interpreted on this arm’s terms. The five sum to an arm strength of 0–10, then a female-applicability multiplier is applied.
| Dimension | What it measures here | 0 | 1 | 2 |
|---|---|---|---|---|
| Corroboration | independent corroboration | a single primary study | a single systematic review / meta-analysis, or two independent studies | three+ independent and consistent studies, or one large, well-powered, low-bias RCT |
| Rigor | study design / risk of bias | case report / preclinical | observational or small trial | RCT, meta-analysis, or active guideline |
| Specificity | this drug, this condition | proxy only | drug named, condition adjacent | both named directly |
| Plausibility | mechanism | asserted | plausible | evidenced in relevant biology |
| Consistency | do results agree in direction | conflicting | mostly one way | unanimous (a single study is scored neutral, not penalized) |
How this arm is held in check
A single source caps corroboration at 1. A lone systematic review or meta-analysis is one synthesis, not independent replication, so the trials pooled inside it are not counted as separate sources. Corroboration 2 is reserved for three+ genuinely independent studies or one large, low-bias pivotal trial. This follows the modern evidence view that a single large, well-conducted RCT can outweigh a meta-analysis of small, biased trials (Oxford CEBM Levels of Evidence), with risk of bias judged per the Cochrane Handbook. Imprecision is handled by hard rules, not model judgment:a very small sample (N < 30, or < 300 events) caps corroboration and rigor at 1; an effect with a p-value but no sample size and no confidence interval is flagged precision-unknown and routed for full-text review rather than credited.
Every arm’s score is discounted by how far its evidence was generated in women.
This is the correction the rest of drug development skips. Each arm’s 0–10 strength is multiplied by a bounded factor judged on whether that arm’sevidence is in or about women. The same drug can carry a different multiplier in different arms: a women’s-health forum is inherently female (F1), while a male-derived trial sits at F5.
The bands follow the reporting standards that ask for sex to be analyzed, not assumed: the SAGER guidelinesfor sex and gender reporting, and Region Stockholm’s Janusmed Sex and Gender database of clinically relevant sex differences. Two guardrails hold throughout: it discounts, it never excludes (the floor is ×0.50, so a male-derived drug still surfaces, clearly marked), and absence is not inferiority. Only F6 means a known disadvantage (the kind the zolpidem dose reduction eventually formalized); F4 and F5 simply mean “not yet shown in women.”
| Band | What the evidence shows | Multiplier |
|---|---|---|
| F1 · Female-generated | female-specific condition, or studied in women / ≥80% female | ×1.00 |
| F2 · Represented & equivalent | ≥50% female and a sex-stratified analysis found no meaningful difference | ×1.00 |
| F3 · Represented, not analyzed | ≥50% female but results not broken out by sex | ×0.90 |
| F4 · Underrepresented / extrapolated | <50% female, or mixed with no sex analysis | ×0.75 |
| F5 · Male-derived / female-excluded | <30% female or male-only, applied to a female context | ×0.60 |
| F6 · Sex-dependent disadvantage | verified evidence the drug behaves worse in women | ×0.50 ⚠ |
Arm strength × multiplier determines the tier.
The final arm score (strength 0–10 × applicability multiplier) sorts into four tiers. The cutoffs were set on reason, then frozen against the real score distribution and hand-judged boundary pairs, so they carve the board where the confidence boundaries actually fall.
guideline-grade: independently replicated, low-bias, in women
good evidence, not yet definitive
a real early lead worth watching
thin or single-source; surfaced with heavy caveat
Contradictions in the underlying evidence cap the consistency score and are shown explicitly rather than averaged away, following the Cochrane and CEBM view that disagreement is meaningful information.
Cross-condition is a derived lens.
An earlier version of Whel scored cross-condition signals as a fourth evidence arm. It was demoted because it is inferential: no source directly says “this drug treats condition Y,” so scoring it on the same observed-evidence dimensions would dress up a prediction as evidence. Cross-condition reasoning still matters, since the six conditions converge on shared biology (estrogen signaling, inflammation, metabolic regulation, pain processing), but it now sits on top of the three evidence arms as a clearly-labelled derived-hypotheses layer, never blended into a pair’s score.