Your VCF, checked against its own header before anyone trusts it

Paste a VCF — the header and as many records as you have. A reader in this page parses every ##INFO, ##FORMAT and ##contig declaration, validates each record against them, pins the reference build from the declared contig lengths, and computes Ti/Tv, missingness and per-sample depth. Then the model gives you the verdict, the findings and an ordered bcftools filter plan.

Both examples ship with a saved model run for every lane, so you can see all three complete outputs without signing in and without spending a credit.

Your VCF nothing pasted yet

Drag a .vcf in, or Everything is read in your browser. Nothing uploads until you run. A gzipped .vcf.gz must be decompressed first — this page reads text.

What kind of callset this is

This changes which QC expectations apply — a somatic callset is not judged against germline Ti/Tv or het/hom ratios.

Context — optional, but it changes the answer

Paste a VCF to price the run.

verdict

reference build

caller

records

samples

findings

What the numbers say

The model's reading of the metrics this page computed. A not-applicable row is a metric that does not apply to this kind of callset.

Metric	Value	Expected	Reading	Note

Checks

Check	Status	Evidence	Requirement

Findings

Filter plan

Annotation state

Genes the file names

Only gene symbols present as text in the pasted file. If your VCF is unannotated this is empty by design — nothing here is recalled from memory.

Shortlist

#	Locus	REF>ALT	Gene	Consequence	Why	Confidence

Lookup plan

The queries that would establish what this page refuses to assert. Read each command before running it — nothing here is executed by this page.

Evidence gaps

Ingestion readiness

Requirement	Status	Evidence	Remediation

Blockers

Attribute schema

Attribute	From	Type	Note

Ingest plan

Cohort risks

Prescan reconciliation

Every flag this page raised, and what the model did with it. A flag the model never mentioned is called out here rather than quietly dropped.

Summary

Raw model reply

The reply did not parse as the structured review, so it is shown as it arrived.

What this does, and what it does not

The prescan is a real VCF reader. Every ## meta line is parsed into its structured <ID=…,Number=…,Type=…> form with quoted Description values split correctly, so a comma inside a description does not corrupt the declaration. Each record is then validated against what the header declared: an INFO or FORMAT value whose comma count disagrees with its Number — including the A, R and G forms — is flagged with the line it happened on, because a reader that trusts Number will read past the end of that field into the next one. The reference build is pinned by matching declared contig lengths against known assemblies rather than guessed from the file name, so chr1 at 248,956,422 bases is GRCh38 and at 249,250,621 is GRCh37 — the distinction that silently corrupts a merged cohort.

It reads; it does not call variants, it does not fetch anything, and it never reaches your reference or your BAMs. Everything it finds is handed to the model as facts with stable ids that must each be reconciled, and the reconciliation is checked here afterwards — if the model skips a flag, this page says so. The interpret lane is deliberately constrained: it may only name genes that appear as text in your file, and every claim that would need a database becomes a gget query instead of an assertion. That is a real limit, and stating it is more useful than a confident sentence about gene function that nothing in your file supports.

Nothing to hand? Load the , a single-sample file with undeclared INFO keys, an unsorted block, a mixed contig naming convention and a low Ti/Tv, or the , a joint call carrying VEP consequences, a high-missingness sample and multi-allelic sites. Both replay a saved run for every lane, for free.