Counting rules (v0.1)

Headline — public experiments matched to a standard line ID
  Count a public metadata record once when it has at least one match with
  status "accepted" and confidence >= 0.75 to a human
  pluripotent cell line (Cellosaurus category Embryonic stem cell or Induced
  pluripotent stem cell, NCBI taxonomy 9606, or an hPSCreg name attached to
  such a line).

  Grain. One record is one BioSample, GEO sample (GSM), SRA experiment
  (SRX/ERX, not each run), ArrayExpress SDRF sample, CELLxGENE dataset, or
  ENCODE accession cited by Cellosaurus. Cells are not counted. Raw reads are
  not stored.

  Deduplicate. Records that share a BioSample accession count once. A GEO
  series (GSE) does not count when an accepted sample from that series is
  already in the index. Records that cannot be linked stay separate; a GSM
  and a SAMN for the same biology can still both count when the crosswalk is
  missing. That gap is reported rather than guessed.

  Do not count. Flag non-human lines (any species other than human), organism conflicts, ambiguous short names
  (two human pluripotent lines share the alias and nothing longer breaks the
  tie), low-confidence hits, and non-pluripotent stem populations
  (mesenchymal, hematopoietic, cancer stem, and similar) are stored with a
  status and excluded from the headline. Skip controlled-access records (dbGaP,
  EGA, non-public consent) are not ingested.

Distinct lines
  Distinct Cellosaurus accessions (CVCL_) among accepted matches. hPSCreg
  names are an alias of the CVCL line when Cellosaurus records the cross-reference.

Linked trials
  ClinicalTrials.gov studies returned by the iPSC/ESC/hPSC query that the
  matcher links to a human pluripotent line at confidence >= 0.85, using
  distinctive names only (no bare H1/H9/H7). The broader hPSC trial list is
  reported separately and is not added to the experiment headline.

Matcher 0.2 accepts a short name only for the WiCell codes H1, H7, H9, H13,
and H14, and only when the cell-line field is exactly that code or the text
has pluripotent context. Clone numbers, control labels, and type-word stubs
(ON-iPSC reduced to "on") are not matches. A name inside a longer hyphenated
code (SH-iPS15-10) is not a match. Differentiated progeny still count when
the parent line is named. Matcher 0.2.1 treats a filled non-human organism
name (Sus scrofa, chimpanzee, E. coli, mixed sample) as a conflict even when
tax_id is blank. Matcher 0.2.2 does not accept a title that names two lines,
and it does not let a title override a different line in the cell-line field.
An unresolved short name in that field (bare H1, which is shared) still blocks
a different line named only in the title. A name that is only stopwords plus
a number (AND-3 read as "and 3") counts only when it is the whole field.

Reviewed precision is an agent read of a 500-row sample, stored separately
from human_label. It is not a human audit. The stats page states the sample
size, the reviewed precision, and the Wilson interval from that read.