Methodology

How what you read here is produced, what the labels mean, and what they do not mean.

The circuit

The system discovers scientific literature in PubMed and Europe PMC, normalizes it, reads it and extracts statements. Each statement is linked to the paper supporting it and to the verbatim excerpt behind it.

Between automated reading and publication there are three gates. None can be skipped.

The three gates

1. Reference validation. No AI involved. The system resolves every DOI against Crossref and every PMID against NCBI, and compares what the registry returns with the paper we think we hold: an identifier that exists but points at a different work does not pass. It then checks the excerpt appears verbatim in the text we hold. A fabricated reference is not a warning: it is a failure that stops publication.

2. Adversarial review. A second agent tries to defeat each statement: does the source really say this? in which species? how many animals? was there a control group? is the dose achievable? It may lower an evidence level, never raise it.

3. Human review. Required for anything touching therapeutic efficacy, treatment or cure. Those statements are not published as conclusions until a person has read them.

What “verified excerpt” means

It means one thing only: that this text appears verbatim in the cited paper. We checked it character by character against the stored document.

It does not mean the statement is true, established, or that the finding holds. A study can be badly designed and its quotation still be exact. Judging strength is what the evidence level is for, not the excerpt.

What confidence means

The confidence figure (0–100) answers one question: how faithfully does this statement represent what the source says?

It is not the probability that a treatment works, nor that the finding is correct. A confidence of 100 about a study of five cats with no control group is still a study of five cats with no control group.

Evidence levels

They measure proximity to the cat, which is where FeLV research most often overreaches.

A — controlled clinical study in cats, adequate sample, replicated.
B — direct experimental evidence: feline in vivo, or FeLV-specific in vitro with controls.
C — preliminary: small sample, no control group, single case, unreplicated.
D — indirect: another species, another retrovirus, mechanistic analogy.
E — computational, no experimental confirmation.

Caps are applied automatically: a study not in cats cannot exceed D; in vitro only cannot exceed B; no control group cannot exceed C.

Types of statement

Evidence — the source states this directly.
Inference — reasoned from the source, which does not say it.
Hypothesis — a proposal; nothing asserts it yet.
Open question — the analysed literature does not answer this.

Findings about other species or other retroviruses (HIV, HTLV) are inference when applied to cats, never evidence, unless the source itself makes the feline claim.

What we analyse and what we do not

We analyse with AI only open-access papers whose licence permits it. Many publishers expressly reserve text-and-data-mining rights and we respect that reservation.

Papers that are not open access are indexed by their bibliographic data and linked to the source, but neither processed nor quoted. We never circumvent paywalls.

Artificial intelligence is not a source

Models interpret; they do not support. They can be wrong, which is why every statement is checkable against the original paper: the DOI is always in view.

Each page shows which agent produced it, on which model, and when. That traceability exists so you can check us, not to advertise automation.

Errors and corrections

This system gets things wrong. When it does, we want to know.

If you spot a mistake, write to us in Voices naming the page and, if you can, the source that shows it. A person reviews it and, if you are right, the page is corrected and the correction published.

Knowledge is never silently overwritten: a statement that changes produces a new version, and the previous one remains readable.

What this project does not do

It does not diagnose, prescribe, recommend doses, or replace a veterinarian. A research hypothesis is not a treatment, however promising it sounds.

If your cat is ill or has tested positive for FeLV, see a vet.