Laboratory evidence that is public, and effectively invisible.
This is the one result on this site with nothing to do with my variant. It is a measurement of a database, and it applies to every gene in it.
When a laboratory measures what a genetic variant actually does to a protein, it can deposit that measurement in ClinVar, the public database clinicians and genetic counsellors rely on. That evidence is real, it is expensive to produce, and it is filed correctly. It is also absent from the compact version of the record that automated tools request first, and none of the tools surveyed reads it from the full version either. Several already have the full record on disk and simply never look at the field.
The scan
A single pass over one release, 4,531,457 records, on one machine. For a single variant the compact record and the full record differ by an order of magnitude in size, and the functional effect, the numeric result, the severity call and the evidence code exist only in the larger one. The compact format does not declare a functional element at all.
One correction has to travel with any count.
One laboratory accounts for the overwhelming majority of records carrying this evidence, through a single very large bulk submission. Quoting the raw total without excluding it would treat one submission as hundreds of thousands of independent characterisations. Every version of this analysis excludes that depositor before reporting anything, and it is stated first because it is the most likely way the result could mislead.
Why unclassified does not mean neglected
A classification in ClinVar requires that a variant has been seen in a person and submitted with clinical context. Functional evidence alone does not trigger one, and it is not supposed to. So many of these are unclassified because no carrier has been reported yet. The evidence has been banked ahead of the first patient, sitting in a field the ordinary programmatic route does not return. Whether it is used when someone does turn up depends on whether the person interpreting the variant knows to look in the full record.
Two variants in the same small region of the gene I was studying carry the strongest functional evidence tier the framework allows, and have never been classified by any submitter. Both are absent from population data. Noticing that is what prompted this census, and the census then showed it is not a peculiarity of that region or that gene.
These numbers were wrong once, and the correction is public. An earlier build of the data file reported a larger pool, because the exclusion filter was tested against a shortened copy of each record's submitter list rather than the real one. That accounted for 911 records exactly, and a second defect readmitted 15 more. The published numbers never depended on the faulty file, and the full working is kept as a companion note rather than the file being swapped out quietly.
Detection is by field presence, not by quality. A record counts if it carries the block. Whether every block holds a meaningful measurement was not assessed. The evidence code count comes from pattern matching over free text with inconsistent capitalisation, so it is a floor rather than a total. One release, one snapshot: these counts drift weekly.