About

Reading a variant across registries means opening ClinVar, then gnomAD, then a literature search, and reconciling three different identifiers for the same thing by hand. VarInsight does the resolving and the fetching, and then gets out of the way.

The constraints it is built under

Every assertion carries a link to the record it came from. Nothing is summarised, re-worded or scored. There is no account, no database, and no analytics. There is no language model anywhere in it — the tool has no opinion to offer and no mechanism for forming one.

Those are not incidental. A tool that displays other people's assertions, with attribution, and generates none of its own is a different kind of thing from one that interprets — in how much you should trust it, and in what it is permitted to be.

One section does arithmetic on what the registries returned: the ACMG evidence codes apply published thresholds to a frequency, a predictor score and a consequence, each row naming the publication behind it. It stops there. It reports six of the guideline's twenty-eight criteria, lists the other twenty-two with what each would take, and never combines them into a classification — that step is the curation, and it belongs to a qualified person.

Plain English, without paraphrasing

Because nothing is re-worded, a result is written in the registries' own terms —LOEUF, missense_variant, criteria provided, multiple submitters, no conflicts. Every one of them is underlined and carries a plain-English definition, in place or on hover, linked to the published definition it restates. See the glossary. A definition explains a word; it is never a statement about the variant on screen, which would be the interpretation this tool exists not to make.

How it works

A static site with one script. What you type is parsed into a canonical form, and then every registry is queried directly from your browser — none of them through a copy or an aggregator. They are independent: if Europe PMC is down you still get the variant, and if dbSNP is down you still get the literature. Each section says which registry it came from and whether that registry had nothing, which is a different thing from an error.

Sections that need something another one supplied — an allele frequency needs a position, a trait association needs an rsID — wait for that and nothing else. Waiting for a key is not the same as one registry standing in front of another, and no answer is ever put together from a name that merely looks similar.

The genome build

Shown on every result, because the same variant sits at different coordinates on GRCh37 and GRCh38 and both are real positions. The build is requested explicitly on every call and then checked against the record that comes back; a record in the wrong build is refused rather than displayed.

What it will not become

No batch upload, no VCF parsing, no API, no accounts, no summarisation. Those are all reasonable products; they are not this one.