GenomeLens is run end-to-end against the Genome in a Bottle Ashkenazim trio — HG002, HG003, HG004 — using the publicly released GIAB GRCh38 benchmark callsets and their defined confident regions.
526 / 526 explicit benchmarked calls concordant
0 mismatches
0 normalization failures
Per sample: HG002 169, HG003 187, HG004 170.
Sample ID — HG002 / NA24385 (son), GIAB Ashkenazim trio.
Benchmarked calls — 169, all concordant. Zero mismatches, zero normalization failures.
Truth and callability scope — the GIAB GRCh38 benchmark callset for this sample, restricted to its defined confident regions.
The 169 counts explicit calls compared within the GIAB benchmark scope — positions where the truth set states a genotype. Positions the truth set declines to state are reported as callability, not as agreement.
Sample ID — HG003 / NA24149 (father), GIAB Ashkenazim trio.
Benchmarked calls — 187, all concordant. Zero mismatches, zero normalization failures.
Truth and callability scope — the GIAB GRCh38 benchmark callset for this sample, restricted to its defined confident regions.
The 187 counts explicit calls compared within the GIAB benchmark scope — positions where the truth set states a genotype. Positions the truth set declines to state are reported as callability, not as agreement.
Sample ID — HG004 / NA24143 (mother), GIAB Ashkenazim trio.
Benchmarked calls — 170, all concordant. Zero mismatches, zero normalization failures.
Truth and callability scope — the GIAB GRCh38 benchmark callset for this sample, restricted to its defined confident regions.
The 170 counts explicit calls compared within the GIAB benchmark scope — positions where the truth set states a genotype. Positions the truth set declines to state are reported as callability, not as agreement.
Per-sample counts sum to the 526 reported above (169 + 187 + 170).
The figure counts positions where the truth set states a genotype. Positions the truth set declines to state are not counted as agreement — they are reported separately as a callability result. Counting them toward concordance would inflate it.
This matters more than the percentage. A concordance rate whose denominator is unpublished is not a measurement.
Precision. 526 comparisons with zero errors is not evidence of a 100% rate. By the rule of three, the one-sided 95% lower bound is approximately 99.4% pooled, and approximately 98.3% at the per-sample counts of roughly 170. The point estimate is reported with that bound in mind.
Direction. The figure measures agreement where both the truth set states a genotype and the pipeline reports one. It is a genotype-concordance measure, not a recall measure: a position where the truth set states a genotype and the pipeline emits nothing is a separate quantity, and 0 mismatches cannot detect it.
Independence. HG002 is the son of HG003 and HG004. Roughly half of HG002's alleles are present in a parent by descent, so the 526 comparisons are not 526 independent observations, and the trio is one nuclear family of a single ancestry rather than a population sample.
What the benchmark is not: not whole-genome accuracy; not clinical validation; not an endorsement by NIST or GIAB. Technical concordance says a genotype matches an external truth call within a defined scope. It says nothing about clinical utility.
Chromosome scope. The GIAB small-variant benchmark used here covers the autosomes, chr1–22. chrX, chrY and chrM therefore have zero benchmark coverage — a different statement from the modality limits in LIMITATIONS.md, which concern what a variant callset can resolve at all. Nothing here is evidence about sex chromosomes or mitochondrial variants.
Regions outside the confident-region definition; variant classes the truth set does not adjudicate; and any capability requiring read-level evidence rather than a variant callset. These are stated rather than silently folded into a pass.
Published artifacts carry the engine build that produced them, the input callset and truth-set version, and the region definitions applied. A result that cannot be attributed to a specific build is not publishable as evidence.
This release does not yet meet that standard for the benchmark figure. The truth-set release version, confident-region file identity, engine build identifier and retrieval date are not published alongside the 526 figure, and the rule by which the benchmarked positions were selected is not stated. Until they are, the figure should be read as a reported result rather than as an independently re-derivable one. See RELEASE-STATUS.md.
A synthetic whole-genome variant-call cohort is designed to test behavior at scale: failure modes, output distributions, economic coverage and systems behavior across many inputs.
It is explicitly planned and not yet executed, and when it runs it will answer a different question from the GIAB benchmark. Synthetic inputs test stability and coherence at scale; they are not evidence of real-world accuracy, disease prevalence, or clinical validity, and they do not substitute for real-genome validation.
It is not run in this release, and is not described as existing.