Methodology

Varia is post-test interpretation software that reads user-supplied raw genotype output. The variant calls Varia interprets were generated upstream by other testing systems (consumer direct-to-consumer arrays, CLIA- or CAP-accredited whole-genome sequencing, or targeted clinical panels). Varia does not perform genotyping, sequencing, or any in vitro molecular detection.

You load a DNA file from your device and Varia returns curated findings drawn from peer-reviewed scientific literature and named institutional authorities. The product runs entirely in your browser. Your DNA file is never uploaded.

Applicable regulatory frameworks include the FDA General Wellness final guidance (January 6, 2026 reissue), 21 USC 360j(o) (Cures Act software exclusions for displaying medical information), and 21 CFR 880.6315 (general wellness; low-risk devices) where applicable. Varia is not a Genetic Health Risk Assessment System under 21 CFR 866.5950 because Varia performs no variant detection.

Three editorial commitments shape every finding Varia displays:

Curated. Varia's current archive represents 49 carefully chosen variants across twelve domains. By contrast, Promethease ships 111,702+ entries unfiltered and 23andMe reports against an opaque internal catalog. Varia picks the smaller set deliberately, because each variant that we report on has been through a seven-pass editorial review and survives a documented evidence threshold.

Transparently sourced. Every finding cites its primary literature. Every authority Varia consults appears on the Institutional Authority References page with the date Varia last reviewed it. Every detector that watches for upstream changes (ClinVar, ClinPGx, ClinGen, PharmGKB, PharmVar, Retraction Watch, ACMG SF) runs on a documented schedule with results committed to the editorial event log in our public repository.

Editorially disciplined. Varia distinguishes findings that come with clinical authority (ACMG/AMP classifications, CPIC pharmacogenomic guidelines) from findings that come from Varia's own editorial assessment (common variants and lifestyle traits where no clinical authority exists). The framework is explicit. The reasoning is visible. The cadence at which findings get reviewed is public.

Varia is not a substitute for clinical genetic testing or genetic counseling. The recommendation boundary is bright: Varia surfaces variant names and biomarker or drug class names. It does not recommend doses, supplements, products, or protocols.

Varia is post-test interpretation software. It reads a DNA file you already have from another testing system (a consumer chip export, a clinical lab, or whole-genome sequencing). Varia does not run a new genetic test and does not perform sequencing in a lab.

You open your DNA file from your device, and Varia analyzes it directly in your web browser. Your file never leaves your device.

Varia follows FDA General Wellness guidance (reissued January 2026), the Cures Act exclusion for software that displays medical information (21 USC 360j(o)), and general-wellness device policy (21 CFR 880.6315) where those frameworks apply.

Three editorial commitments shape every finding Varia displays:

Curated. Varia reports on 49 variants chosen with care, organized across twelve health-relevant categories. Other consumer-genomics products either show you every variant they find without filtering (Promethease shows over 111,000), or report on a much larger set without telling you why they picked those (23andMe). Varia keeps the list intentionally short because every variant has been reviewed through a seven-step quality process before earning a spot in the catalog.

Transparently sourced. Every finding Varia shows you links back to the scientific papers it draws from. The major medical authorities Varia consults (the same authorities your clinician uses) are listed openly with the date each was last reviewed. Computer programs check those authorities for updates on a published schedule, and every change gets logged in a public record so anyone can see what changed and when.

Editorially disciplined. When a clinical authority has already graded a variant (like for rare-disease risks or medication response), Varia shows you their grade directly. For variants where no clinical authority weighs in (common health-trait variants, lifestyle traits), Varia applies its own grading and shows the reasoning behind it. Either way, you see how the grade was reached and when it was last reviewed.

Varia is not a substitute for clinical genetic testing or working with a genetic counselor. The line Varia holds is clear: Varia tells you what variants you have and what published science says about them. Varia does not recommend specific doses of medication, supplements you should take, or lifestyle protocols you should follow.

Varia tracks 49 variants across twelve domains: APOE Alzheimer's, chromosome 19 corridor, Alzheimer's clearance pathways, cardiovascular lipid metabolism, neuro-pharmacogenomics, methylation and detoxification, hormonal androgen, longevity and telomere biology, inflammation, metabolic diabetes, physical performance, and the APOC1 haplotype.

The catalog scope is intentional. Over the last two decades, genomic interpretation has been a bottleneck. Sequencing throughput scaled roughly six orders of magnitude in the same period, but the functional and clinical interpretation of variants grows linearly through peer-reviewed publication. Most of the human genome remains uninterpreted at the clinical-actionability level. A consumer-genomics product that surfaces every variant present in your file either oversells the actionability of those findings or hides the most defensible findings inside an unmanageable list.

Varia chooses the smaller, defensible list. Every variant in the catalog passes one of two qualification paths:

• Authority-backed: the variant has an ACMG/AMP classification in ClinVar (Pathogenic, Likely Pathogenic, or VUS) with documented review status, OR a CPIC pharmacogenomic guideline at evidence level A or B, OR a PharmGKB clinical annotation at level 1A or 1B. For these variants, Varia surfaces the institutional grade verbatim and adds a confidence chip indicating ClinVar review status.
• Varia-graded: the variant is a common-trait risk modifier (lipid metabolism, methylation pathway efficiency, inflammation response) or lifestyle association. No clinical authority grades these. Varia applies its own evidence assessment per the five-axis framework below and labels the grade transparently with citations behind it.

HLA loci (HLA-B*15:02, HLA-B*57:01, HLA-DQ2 / HLA-DQ8 and similar) are deferred from the V1 catalog. Consumer chip imputation is unreliable for HLA, and whole-genome calling requires specialized pipelines beyond Varia's current scanner. When HLA returns, it returns with whole-genome-only restriction.

Varia tracks 49 variants across twelve categories: APOE Alzheimer's risk, the chromosome 19 region surrounding APOE, Alzheimer's-related clearance pathways, cardiovascular and cholesterol metabolism, medication response (how your body processes drugs), methylation and detoxification, hormonal and androgen pathways, longevity and aging biology, inflammation, metabolic and diabetes-related variants, physical performance traits, and the APOC1 haplotype.

The catalog size is intentional. Over the last two decades, our ability to sequence DNA has grown roughly a million-fold, but our scientific understanding of what each variant actually does has only grown step by step through peer-reviewed publication. Most of the human genome remains scientifically uninterpreted at a level that supports actionable medical decisions. A consumer-genomics product that shows you every variant in your file is either overselling what those findings can tell you, or burying the strongly-supported findings inside an unmanageably long list.

Varia chooses the smaller, defensible list. Every variant in the catalog earns its spot through one of two paths:

• Authority-backed: a recognized medical authority has already published a grade for this variant. ACMG and ClinVar grade rare-disease risk variants. CPIC and PharmGKB grade medication response variants. When that authoritative grade exists, Varia shows it to you directly with a chip indicating how strong the consensus is.
• Varia-graded: the variant is a common health-trait or lifestyle variant. No clinical authority grades these kinds of variants. Varia applies its own five-axis evidence framework (described below) and shows the reasoning publicly.

HLA loci (a family of immune-system variants linked to drug allergies and some autoimmune conditions) are not in the V1 catalog. Consumer DNA chips can't reliably read these regions, and reading them from whole-genome sequencing requires specialized analysis software Varia doesn't yet run. When HLA returns, it will be available only for whole-genome sequencing files.

Each finding card in Varia carries five axes of structured grading.

Variant Class identifies what kind of finding this is. Four possibilities: Rare pathogenic variant (Mendelian disease association in ACMG/AMP framework), Pharmacogenomic (drug response with CPIC guidance), Common variant: risk modifier (GWAS-style association with a complex trait), and Lifestyle trait (athletic performance, caffeine metabolism, similar).

Strength applies only to Varia-graded findings (risk modifier and lifestyle classes). Five labels: Well-replicated (multiple independent Tier 1 publications, consistent effect direction, large total cohort size), Consistent (multiple cohorts with consistent effect), Mixed evidence (effect direction varies across studies), Single-study (one strong study; replication pending), and Disputed by recent work (recent literature challenges the prior consensus). Authority-backed findings (Pathogenic, Pharmacogenomic) do not carry a Varia Strength label; the ACMG or CPIC grade controls instead.

Conversation Context (VEGS § 4 Axis 2) describes how institutional authorities relate to the finding. Four labels: Authority-graded; medication-relevant; Authority-graded; condition-associated; Guideline-referenced; physician-discussion-common; and Educational; no authority framework.

Conversation Priority (VEGS § 4 Axis 2b) is Varia's editorial framing of discussion warranting. It is not a clinical urgency or triage signal. Four labels: Institutionally flagged; Routine discussion; Background context; and Evidence insufficient to characterize.

Input Confidence describes how Varia interprets the type of data your file represents. Eight buckets ranging from high-coverage whole-genome sequencing (WGS) with verified lab provenance down to consumer chip exports. Varia determines this automatically by inspecting your file and confirming it with you at scan time.

Input Provenance Confidence distinguishes sequencing-quality metrics from origin certainty. High-depth metrics alone do not establish that a file originated from a verified sequencing pipeline; reconstructed, transformed, or fabricated VCFs can satisfy quality heuristics while obscuring origin. Three buckets: verified sequencing origin (cryptographically signed laboratory export or verified provider metadata), probable sequencing origin (header signatures and quality metrics match known sequencing laboratories), and unverified origin (file lacks provenance signals; quality may still pass, but origin is not establishable).

The five axes work together. Authority-backed Pathogenic findings display clinical-grade chrome (ACMG classification + ClinVar Confidence Chip + ClinGen Actionability score + link to ClinVar) only when Input Confidence is high-coverage WGS and Provenance is verified or probable. Pathogenic findings on chip-grade input route to a safety surface card pointing toward CLIA-certified clinical confirmatory testing rather than displaying the clinical-grade chrome on data that cannot support it.

This routing is the architectural answer to a known failure mode of consumer genomics. A 2018 study (Tandy-Connor et al., Genetics in Medicine, 20:1515-1521) found that approximately 40% of variants reported in cancer-predisposition genes from consumer chip-based DNA tests were false positives on confirmatory clinical testing. Most consumer-genomics products display chip-derived Pathogenic findings with the same visual chrome as findings from clinical sequencing. Varia separates them deliberately.

Each finding card in Varia shows you five different pieces of information, working together to communicate confidence and clinical relevance.

Variant Class tells you what kind of variant you're looking at. There are four possibilities: a rare variant linked to a disease (the kind a clinical geneticist would investigate), a variant affecting how your body processes medication, a common variant that modifies your risk for a complex trait (like cholesterol levels or inflammation response), or a lifestyle-trait variant (like athletic performance or caffeine processing).

Strength is Varia's grade of how well-supported the science is. This applies only to the Varia-graded variants (common-trait and lifestyle); for authority-backed variants, the authority's grade controls instead. Five labels: Well-replicated (consistent findings across multiple major studies and large populations), Consistent (findings agree across multiple studies but the evidence base is smaller), Mixed evidence (different studies find different things), Single-study (one strong study; we're waiting for replication), or Disputed by recent work (newer research has challenged what older studies found).

Conversation Context states which authorities, if any, have graded or referenced the finding (medication guidelines, disease classifications, general guidelines, or research-only context).

Conversation Priority is Varia's editorial label for how the finding fits into a conversation with a clinician. It is not an urgency score. Four labels: Institutionally flagged; Routine discussion; Background context; and Evidence insufficient to characterize.

Input Confidence tells you how confident Varia is in the type of data your file represents. Eight categories ranging from high-coverage whole-genome sequencing (the gold standard) down to consumer DNA chips. Varia figures this out automatically by inspecting your file and confirming with you at scan time.

Input Provenance Confidence is separate from how technically good your data is. It addresses where your data came from. Files that look high-quality but lack verifiable lab origin signals can sometimes be reconstructed or transformed in ways that hide their actual source. Three categories: verified lab origin (a signed export or verified metadata bundle from a sequencing lab), probable lab origin (the file's quality metrics and header signatures match what a real sequencing lab would produce), and unverified origin (the file might be high-quality, but we can't confirm where it came from).

The five axes work together. For example, a finding linked to a rare disease only displays with full clinical-grade visual confidence when your input is high-coverage whole-genome sequencing from a verified or probable lab origin. Rare-disease findings from consumer-chip data display differently: rather than showing clinical-grade chrome that the data can't actually support, Varia displays a safety card pointing you toward confirmatory testing at a clinical laboratory.

This careful separation is the architectural answer to a known problem in consumer genomics. A 2018 study (Tandy-Connor et al., published in the journal Genetics in Medicine) found that approximately 40% of variants reported in cancer-predisposition genes from consumer DNA tests turned out to be false positives when re-tested in a clinical laboratory. Most consumer-genomics products show the same visual confidence on chip-derived findings as they do on clinical sequencing. Varia separates the two deliberately because the data quality is meaningfully different.

VEGS v1.1 splits the former combined "conversation" axis into two independent axes per § 4. Conversation Context is descriptive: it states which institutional authorities cover the finding. Conversation Priority is Varia editorial framing: it states how the finding warrants discussion language. Neither axis recommends treatment, screening, or medication changes.

Conversation Context (Axis 2)	Conversation Priority (Axis 2b)
Authority-graded; medication-relevant	Institutionally flagged
Authority-graded; condition-associated	Routine discussion
Guideline-referenced; physician-discussion-common	Background context
Educational; no authority framework	Evidence insufficient to characterize

The table lists the four allowed values on each axis. They are not paired one-to-one; graders assign each axis independently per VEGS § 5 routing. See the Editorial Standards page for authority-surfacing rules.

Varia uses two separate labels. Conversation Context answers "does a medical authority already grade this?" Conversation Priority answers "how does Varia frame discussion of this finding?" Neither label tells you what treatment to start.

Conversation Context	Conversation Priority
Authority-graded; medication-relevant	Institutionally flagged
Authority-graded; condition-associated	Routine discussion
Guideline-referenced; physician-discussion-common	Background context
Educational; no authority framework	Evidence insufficient to characterize

The columns list the four options for each label. Varia picks Context and Priority separately for every finding; reading across a row is not how the labels combine.

Per VEGS § 11 and § 11.6: Varia runs a transparent, version-controlled curation pipeline that applies published algorithmic scoring criteria at scale. Grading rules, thresholds, and source policy are human-authored. Software applies those rules mechanically. Residual human editorial steps are documented, quantified, and auditable (see quantified counts below and editorial event log). Grader outputs use an explicit disposition: algorithm_default_applied (conservative default stands; logged, not a standing queue) vs review_required (folds into the monthly review-and-approve surface). PharmGKB publishes a measured manual-override rate (~1.9%); Varia follows the same disclosure posture on review_required resolution only.

Quantified residual human contribution (2026-05-28 close artifacts; disposition reframe 2026-05-29):

• 32 PMID corrections editorially adjudicated (111 locator patches applied; Sprints 9 to 11).
• 15 conflict-of-evidence hand-authored summaries (24 catalog rows including genotype mirrors; Sprint 47).
• 12 CPIC retiering genotype rows human-adjudicated via worksheet (Sprint 28).
• 9 ClinVar authority-contradiction reconciliations resolved (1.9% of 470 graded catalog entries; Sprint R2).
• 628 VEAS citation annotations under human-authored slot templates and voice-anchor criteria (VEAS pipeline v0.2).
• 722 legacy grader events reframed to disposition alert classes (bootstrap ambiguous-threshold cases log as algorithm_default_applied; genuine-action cases as review_required).

Seven meta-editorial governance touchpoints (human-authored; each change logs to the editorial event log):

1. Strength rubric thresholds (cohort N, Tier 1 representation, effect-direction agreement).
2. Source allowlist composition (VEGS § 8 journal tiers).
3. Override-trigger rules (nine categories per § 11).
4. Override resolution for review_required only (public editorial_override fields).
5. Voice anchor criteria layer (VEAS § 6.5; five class anchors).
6. Forbidden-word list maintenance (VEAS § 6).
7. Framework version promotion (joint VEGS + VEAS regression per § 11.5).

Claims Varia does not make: "no per-variant human adjudication" (false); "algorithmic grading replaces all human judgment" (false); "Varia's grading is purely objective" (false); "Varia's catalog is reviewed by a clinical geneticist" (false); "Varia's grading has zero subjective inputs" (false); "per-facet concordance has been validated" (false until the pre-registered R7 study completes; design intent only until then).

Varia grades variants with published rules, not hidden case-by-case judgment. Humans write the rules and thresholds; software applies them the same way to every variant. When the rules cannot decide confidently, the case is logged with a clear disposition: either the conservative algorithm default stands, or it folds into Eric's monthly review surface. Varia publishes how much human work remains on genuine review items, the same way PharmGKB publishes its ~1.9% manual-override rate.

Measured human editorial work in the catalog (as of 2026-05-28): 32 PMID fixes reviewed by hand; 15 conflict summaries written for split literature; 12 medication-response retiers adjudicated; 9 authority disagreements reconciled (1.9% of graded entries); 628 citation annotations under human-authored voice rules; bootstrap threshold cases logged as algorithm-default (not a standing queue).

Varia does not claim the software replaces all human judgment, that grades are purely objective, that a clinical geneticist reviewed your catalog entry, that the process has zero subjective choices, or that concordance against ClinGen/ClinVar has already been proven (that validation is planned, not yet run).

Varia does not collapse ACMG pathogenicity, GWAS population risk, CPIC prescribing actionability, and GRADE evidence certainty onto one significance ladder because those scales measure different things and are not commensurable. Merging them into a single tier hard-codes a hidden weighting and creates false equivalence across claim types. GA4GH VA-Spec and ClinGen keep each authority in its native profile; Varia's v1.2 faceted schema (VEGS § 6.1) follows that pattern with four native-scale facets over one shared provenance footer.

Non-equivalence principle: facet labels and legacy significance tiers are not comparable across facets. A population-risk percentile does not outrank a CPIC level; an ACMG Likely Pathogenic call does not fuse with a GWAS odds ratio. Varia stores facet_metadata.non_equivalence_declared: true on every catalog entry and surfaces disagreement between facets where live divergence exists.

Per-facet concordance against source authorities (ClinGen/ClinVar for Facet A, CPIC/PharmGKB for Facet C) is a design target for a pre-registered retrospective study (R7). No concordance numbers are published until that study completes.

Varia does not squeeze every kind of genetic finding into one ranking because the underlying medical scales measure different things. Rare-disease pathogenicity, population risk, medication guidance, and evidence strength use different rules. Combining them into one label would hide those differences and imply a comparison that the science does not support.

What that means for you: labels on different parts of a finding card are not interchangeable. A risk percentile is not the same kind of statement as a medication guideline level. Varia keeps those scales separate on purpose and flags cases where they disagree.

Varia plans a formal check of how well each facet matches the authorities it draws from. Those results are not published yet.

Population-risk and polygenic outputs are ancestry-sensitive. Effect sizes discovered predominantly in European-ancestry cohorts do not transfer as point constants to under-represented ancestries; Martin et al. (2019) and subsequent multi-ancestry GWAS work document systematic attenuation and LD-structure differences. Varia treats transferability as ranges with explicit uncertainty widening, not single constants.

Facet B stores ancestry-calibration metadata (training_data_ancestry, applicability_flag, uncertainty_widening_flag; populated in R3) and tags catalog conflicts as heterogeneity_by_ancestry where literature documents attenuation (for example 9p21 CAD, VKORC1 warfarin dosing, APOE effect-size transfer). Default product behavior widens uncertainty for under-represented ancestries rather than presenting European-calibrated point estimates as universal.

Many genetic risk estimates were discovered in studies where most participants had European ancestry. Those numbers often do not apply the same way to people from other ancestral backgrounds. Varia treats population-risk findings as ancestry-sensitive: effect sizes are shown as ranges with wider uncertainty when the supporting science is less representative of your ancestry, not as one-size-fits-all constants.

When the literature documents ancestry-specific disagreement, Varia labels it openly on the finding (for example when a cardiovascular or medication-response association weakens outside European-ancestry cohorts).

Nine Pathogenic-class catalog entries carry a ClinVar aggregate classification of Likely Benign or Benign while the catalog still lists Pathogenic class (for example PCSK9 rs11591147, LDLR rs6511720, APOB rs693). These entries surface to algorithmic-override review with algorithm-proposed class re-assignment to Risk-modifier.

The user-facing UI shows Conversation Priority Evidence insufficient to characterize with a notice that the entry is pending re-classification review on the public override surface. Rationale and detector triggers are logged in editorial-event-log.md under override trigger authority_contradicts_catalog_class.

Nine rare-disease catalog entries disagree with ClinVar's current benign/likely-benign call. Varia flags them for public review and proposes moving them to the common-variant risk-modifier class. Until review closes, the product shows "Evidence insufficient to characterize" and states that re-classification is pending. Details are in the public editorial event log.

Local-processing means your DNA file is parsed in your browser. The file enters memory from the file picker, gets read locally by JavaScript code Varia ships in the page bundle, and is gone from the session when you close the tab. The browser's network tab will confirm: no file content leaves your device.

Inside that local-processing pipeline, the scanner does what most consumer-genomics tools skip:

rsID-canonical lookup. Variants are matched by rsID first, falling back to chromosomal coordinate only as needed. Most consumer scanners look up by coordinate. That breaks silently when the file is GRCh37 but the catalog is GRCh38 or vice versa, because the same SNP has different coordinates in the two genome builds. Varia normalizes by rsID and confirms the genome build separately.

Build detection from multiple signals. The scanner inspects VCF header fields, chip vendor signatures, and a set of six fingerprint SNPs spanning multiple chromosomes. Older 23andMe v3 chips and chromosome-subset VCFs still detect cleanly because the fingerprint set does not depend on a single locus.

Strand normalization at parse time. Variants reported on the reverse strand by some chip vendors are normalized to the canonical plus-strand representation at parse time. This closes a class of silent corruption where rs429358 reads as the protective allele on one chip vendor but the risk allele on another.

Three-state coverage discipline. Every locus in the catalog reports as one of three states for your scan: TESTED_VARIANT (your file contains a non-reference call at that position), TESTED_REFERENCE (your file confirms reference allele), or NOT_TESTED (the position was absent from your input format). Consumer chip exporters typically conflate "absence" with "negative." Variants-only VCFs invert that. Varia distinguishes them and the UI tells you which is which.

Compound diplotype handling. APOE risk requires phasing two SNPs (rs429358 and rs7412) into the ε2/ε3/ε4 diplotype because the two SNPs combine to determine the protein isoform. Most consumer scanners report the two SNPs separately and ask you to phase them mentally. Varia handles the phasing in the scanner and reports the diplotype directly.

Regression-frozen accuracy. Every commit that modifies the scanner runs against twenty-two named edge-case fixtures spanning different genome builds, sequencing technologies, multi-allelic sites, indels, phased GT, multi-sample VCFs, and chip exports from multiple vendors. If any byte of scanner output changes without explicit fixture acknowledgment, continuous integration blocks the commit. This is the engineering discipline behind the editorial claim that the scanner reads files correctly.

The Varia scanner runs entirely inside your web browser. When you select your DNA file, the file enters your browser's memory, gets read locally by code Varia ships in the page, and is gone from the session when you close the tab. If you open your browser's developer tools and watch the network tab during a scan, you'll see that no file content is uploaded.

Inside that browser-local scan, the scanner does five things that most consumer-genomics tools skip:

It matches variants by reference ID, not by location. Genomes come in two coordinate systems (called builds, named GRCh37 and GRCh38). The same variant has a different address in each. Tools that match by address can confuse one build for the other and silently report the wrong information. Varia matches by reference ID, which doesn't change between builds, so the match is correct regardless of which version your file uses.

It identifies your genome build from multiple signals. Some tools look at one indicator and guess; Varia looks at the file's header, the file format, and a set of six well-known reference variants spread across multiple chromosomes. Older chips and trimmed files that lack one signal still work because Varia doesn't depend on a single source of truth.

It normalizes the way variants are written. DNA is double-stranded, and the same variant can be reported in two different ways depending on which strand is read. Some DNA chips report variants on the reverse strand by convention, which means the same variant can look like the protective version on one chip and the risk version on another if you don't account for the strand. Varia handles this at parse time, so the variant always reads correctly regardless of how the source chip happened to record it.

It tells you the difference between "not found" and "not tested." For every variant in the catalog, Varia reports one of three states for your scan: your file contains the variant, your file confirms reference DNA at that position, or the position wasn't included in your file's coverage at all. Most consumer scanners conflate "not found" with "negative," which makes it look like you don't have a variant when really the test never looked. Varia is explicit about which state applies to each finding.

It handles compound variants correctly. Some genetic risks come from the combination of two variants together, not from either one alone. APOE Alzheimer's risk is the classic example: two specific variants combine into one of three genetic types (ε2, ε3, or ε4), and the type determines the risk. Most consumer scanners report the two variants separately and ask you to figure out the combination yourself. Varia handles the combination automatically and reports your APOE type directly.

Every change to the scanner is regression-tested before it ships. Varia maintains twenty-two named test files covering different genome builds, sequencing technologies, edge cases, and DNA chip vendor differences. If any byte of scanner output changes without explicit acknowledgment, the automated system blocks the change from shipping. This is the engineering discipline behind the editorial claim that Varia reads your file correctly.

Several Varia rules deliberately suppress findings that consumer chip data cannot reliably support.

CYP2D6, CYP2A6, CYP2B6, CYP2E1, UGT2B17, GSTM1, and GSTT1 phenotype calling is suppressed on chip input. These pharmacogenomic genes involve pseudogene homology, copy-number variation, or structural variation that consumer chips cannot reliably resolve. Many products call these genes from chip data anyway. Varia surfaces a suppression card explaining that star-allele or activity-score calling for this gene requires high-coverage whole-genome sequencing with specialized analysis. Whole-genome users who supply input with documented sequencing-laboratory provenance and current CYP2D6-specialized caller versions (BCyrius v1.0.2 or later, Aldy v4 or later, StellarPGx v1.2 or later, or equivalent published ≥95% concordance on the GeT-RM benchmark) get the CYP2D6 authority surface.

DPYD on chip-grade input triggers a positive safety surface. Dihydropyrimidine dehydrogenase deficiency contributes to fluoropyrimidine toxicity in patients receiving 5-fluorouracil, capecitabine, or tegafur. The European Medicines Agency added pre-treatment DPYD screening as a regulatory requirement in April 2020. CPIC published its DPYD guideline in 2017. Chip data cannot reliably exclude DPYD variants. Rather than silently omitting DPYD from chip-input reports, Varia surfaces a safety card pointing toward clinical DPYD testing if a clinician is considering 5-FU, capecitabine, or tegafur for you. This is framed as a clinical-recall pointer to third-party standards of care, not a Varia recommendation.

Pathogenic findings on chip or imputed input route to safety surface. Per Tandy-Connor et al. and the Varia editorial framework, chip-derived Pathogenic findings display with explicit "low-confidence input" labeling and a safety message recommending confirmatory testing in a CLIA-certified clinical laboratory. The ACMG classification chrome is reserved for high-coverage WGS input with verified or probable lab provenance and passing per-variant call quality at that locus.

These suppressions exist because the alternative is showing clinical-grade visual confidence on data that cannot support clinical-grade conclusions. The Tandy-Connor 40% false-positive rate is the empirical anchor.

Some findings can't be reliably reported from consumer DNA chip data. Rather than reporting them anyway and overstating what your file actually shows, Varia surfaces a safety card for these specific cases.

Seven medication-response genes are suppressed on chip data. These genes (CYP2D6, CYP2A6, CYP2B6, CYP2E1, UGT2B17, GSTM1, and GSTT1) involve complex genetic structures that consumer DNA chips can't reliably read: they have look-alike pseudogenes, missing or extra copies in some people, or structural rearrangements that need specialized analysis to call correctly. Many products report on these genes from chip data anyway. Varia shows you a card explaining that calling these genes from your input format isn't supported, and recommends consulting a clinical pharmacogenomics laboratory if your clinician needs this information. If you upload high-coverage whole-genome sequencing data from a verified lab using current specialized software (BCyrius v1.0.2 or later, Aldy v4 or later, StellarPGx v1.2 or later, or a comparable tool with documented accuracy), the CYP2D6 finding becomes available with full authority chrome.

DPYD on chip data triggers a safety message, not silence. A gene called DPYD affects how some patients process certain chemotherapy drugs (5-fluorouracil, capecitabine, tegafur). People with reduced DPYD function can experience severe or fatal toxicity from standard chemotherapy doses. The European Medicines Agency made DPYD pre-treatment testing a regulatory requirement in April 2020. Chip-grade data can't reliably exclude these variants. Rather than silently dropping DPYD from chip-grade reports, Varia surfaces a clear safety card recommending clinical DPYD testing before any treatment with these drugs. This is presented as a pointer to existing standards of care from major regulatory bodies, not as a Varia recommendation.

Rare-disease findings on chip or imputed data route to safety guidance, not clinical chrome. When a variant linked to rare disease is detected in your file but your file is chip-derived or computationally inferred (imputed), Varia labels the finding as low-confidence input and recommends confirmatory testing in a CLIA-certified clinical laboratory. The clinical-grade visual chrome is reserved for findings from high-coverage whole-genome sequencing with verified or probable lab origin and passing quality at that specific position.

These suppressions exist because the alternative is showing clinical-grade visual confidence on data that can't actually support clinical-grade conclusions. The 40% false-positive rate documented in the Tandy-Connor study is the evidence anchor for this caution.

The Varia scanner runs in your browser. Your DNA file enters memory when you choose it from the file picker, gets parsed locally by JavaScript loaded with the page, and is gone from the session when you close the tab. Open your browser's developer tools and watch the network tab during a scan: no file content uploads.

Varia stores a minimal server-side metadata anchor (your email, the timestamp of your most recent scan, the database version at that scan, your unlock status). This anchor exists for recognizing returning users across browser data loss and preserving your unlock status across devices. It does not contain your scan results. Your variants, your genotypes, your interpretations live only in your browser session.

If you close the tab and return later, Varia recognizes your email at the metadata anchor and tells you what has changed in the database since your last scan. You rescan your file (which never left your device) to see the new findings against the updated catalog. The rescan is free.

The Varia scanner runs entirely in your web browser. When you select your DNA file, the file enters your browser's memory, gets parsed locally by code that loads with the page, and is gone from the session when you close the tab. You can open your browser's developer tools and watch the network activity during a scan: no part of your file uploads.

Varia stores a minimal record on its servers (your email, the timestamp of your most recent scan, the version of the database at that scan, your unlock status). This record exists so Varia can recognize you when you return from a different browser or after clearing your cookies. It does not contain your scan results. Your variants, your genetic types, and your interpretations live only in your browser session.

When you close the tab and come back later, Varia recognizes your email and shows you what has changed in its catalog since your last scan. You re-scan your file (which never left your device) to see new findings against the updated catalog. Re-scanning is free.

Whole-genome sequencing (WGS) at 30x average coverage is treated as the highest-confidence input class in Varia. Files from Nebula, Veritas, Dante Labs, Illumina BaseSpace, DNAnexus, and other recognized sequencing laboratories carry header signatures Varia identifies. When those signatures are present and per-variant call quality at each catalog locus passes Varia's documented thresholds (DP ≥ 20 and GQ ≥ 30 for SNVs, DP ≥ 30 and GQ ≥ 40 for small indels, DP ≥ 40 and GQ ≥ 60 for complex regions), the Pathogenic and Pharmacogenomic finding cards display with the full authority chrome.

Variants-only VCFs (high-depth WGS reduced to non-reference positions) receive a clear distinction. Variants-only files contain only positions where your file differs from reference; positions absent from the file are inferred to match reference, not "untested." Varia notes this in the per-card footer: variants-only inputs receive authority chrome with a secondary line explaining the inference.

Low-pass WGS (0.4x to 4x coverage), imputed-from-chip WGS-equivalents, exome panels, and targeted gene panels receive context-aware confidence treatment. Imputed data is treated as chip-equivalent confidence per the editorial framework, regardless of how the file is labeled. Low-pass WGS without imputation falls between chip and high-pass for Pathogenic-class routing.

Whole-genome sequencing (WGS) at 30x average coverage is the highest-confidence input class Varia accepts. Files from major sequencing labs (Nebula, Veritas, Dante Labs, Illumina BaseSpace, DNAnexus, and others) carry signatures in their headers that Varia recognizes. When those signatures appear and the quality metrics at each variant location meet Varia's documented thresholds (the technical numbers are published for transparency: 20 reads minimum and 30 quality score minimum for single-base variants; higher thresholds for harder-to-call variants), rare-disease and medication-response findings display with full clinical-grade visual confidence.

Variants-only VCF files (a common format where the lab includes only the positions where you differ from reference) get a clear distinction in the display. In these files, positions that don't appear are inferred to match reference, not "not tested." Varia notes this in the per-card footer for variants-only inputs.

Low-pass whole-genome sequencing (much lower coverage than 30x), computationally inferred WGS-equivalents (imputed from chip data), exome panels (only the parts of the genome that code for proteins), and targeted gene panels each get appropriate confidence treatment. Imputed data is treated as chip-equivalent regardless of how the file is labeled, because the underlying confidence is the same.

Per VEGS § 10, six detector channels monitor upstream authorities on documented cadence:

• ClinVar XML: weekly (Mondays 00:00 UTC). Primary diff on tracked VCV accessions.
• ClinVar TSV: monthly (first Thursday 00:00 UTC). Review-status reconciliation.
• ClinVar Miner: weekly (Sundays 00:00 UTC). Conflict-report cross-check.
• ClinPGx (CPIC + PharmGKB): weekly API poll (Wednesdays) plus event-driven on GitHub release atom feeds.
• Retraction Watch (Crossref): weekly (Mondays). Every request includes a polite-pool mailto header.
• PharmVar: event-driven via release RSS; monthly fallback poll (first Friday).
• ACMG SF list: event-driven via Genetics in Medicine RSS; annual fallback poll (first Monday of January). Current list: ACMG SF v3.2 (Miller et al. 2023, DOI 10.1016/j.gim.2023.100866).

Per-source SLA targets (detector alert to review-closed; VEGS § 10):

Source	Median target	P95 target
ClinVar XML	≤ 5 days	≤ 10 days
ClinVar TSV	≤ 14 days	≤ 35 days
ClinPGx	≤ 7 days	≤ 14 days
PharmVar	≤ 21 days	≤ 90 days
Retraction Watch	≤ 5 days	≤ 10 days
Retraction Watch (Pathogenic-class)	≤ 24 hours	≤ 24 hours

Pathogenic-class flagged alerts carry a 14-day SLA; a miss auto-downgrades Conversation Priority to Routine discussion pending review. Quarterly SLA metrics publish in the section below. Override soft cap: 3% of catalog OR 3 variants per quarter, whichever is greater (VEGS § 10 metric d).

Varia runs automated checks against ClinVar, ClinPGx (CPIC and PharmGKB data), Crossref retractions, PharmVar releases, and ACMG Secondary Findings updates on the schedules listed above. Results are logged publicly.

Each source has response-time targets (for example, most ClinVar XML alerts should close within five days). Rare-disease alerts must be triaged within fourteen days or the finding's priority downgrades automatically until a human reviews it.

Per VEGS § 10 (M2), Eric's sole recurring editorial cadence is the monthly consolidation loop via scripts/monthly_review/run.py. The tool aggregates new-variant candidates (with PMID pre-verification), open review_required items from the event log, and a draft newsletter section into .sprint-artifacts/monthly-review/YYYY-MM.md. Approved merges run with --apply (database update + npm run build). Newsletter text is draft-only; Varia never auto-sends outbound mail. A missed month fails safe: algorithm defaults already err conservative; unresolved review_required items follow existing SLA auto-downgrade rules.

Once a month, Varia bundles what needs your judgment into one review file: possible new variants, items that truly need a human call, and a draft newsletter you can edit before you send it yourself. Approving candidates updates the catalog in one step.

editorial-event-log.md is the public audit trail for detector alerts, overrides, VEAS generation events, and methodology disclosures. Alert classes include algorithm_default_applied (logged transparency; conservative default stands) and review_required (monthly fold). Legacy override_required rows are append-only reframed, never deleted. Schema v1.1 fields include authority_source, rule_invoked, override_resolution_status, and alert_resolution_criteria.

An alert counts as SLA-closed only when all four resolution criteria are satisfied; auto-close without rationale does not count toward hit rate. Override frequency soft cap per VEGS § 10 applies to review_required resolution volume: 3% of catalog OR 3 variants per quarter, whichever is greater (at V1 scale, the 3-variant floor governs).

Every detector alert and disposition is written to a public log file in the Varia GitHub repository. You can read when something changed, what rule fired, and whether a genuine review is still open. Ambiguous-threshold cases are logged for transparency but do not create a standing queue.

The Citation Library annotates each primary paper cited in the catalog per Varia-Editorial-Annotation-Schema.md (VEAS v0.3). Annotations are generated by the VEAS pipeline (scripts/veas_pipeline/) from PubMed metadata and slot templates (VEAS § 4): study design, sample characteristics, effect summary, replication status, authority alignment, journal tier, publication status, conflict-of-interest disclosure (verbatim), stated limitations (verbatim), and citation-network position.

Voice constraints (VEAS § 6): lay register, mechanism-grounded, no overclaiming language, no first-person outside quotations, no em-dashes. Five voice anchors (§ 6.5) anchor regression. Verification runs four tracks (§ 9): Track 1 schema validation (100% automated, synchronous); Track 2 voice drift (n=10 per model pin); Tracks 3 and 4 spot audits (see below).

Mechanical post-generation sanitization (three transformations only) runs before Track 1: em-dash to comma+space; first-person to third-person; whitespace normalization. Deontic-modal substitution was rejected; deontic hits fail Track 1 and retry or route to override.

Every scientific paper Varia cites can carry a short Varia-written summary built from fixed templates, not free-form prose. The summaries follow strict voice rules and pass automated checks before they appear in the product.

Varia's editorial grading runs through a human-in-the-loop pipeline at scripts/editorial_grader/. See Human-in-the-loop editorial grading, measurement integrity, per-variant scoring trace, editorial event log, and VEAS Citation Library on this page for operational detail. The Editorial Standards page states source-tier and conflict rules.

Cases where the algorithm cannot make a confident determination surface to human override per nine documented override-trigger categories (VEGS § 11). Every override is logged to editorial-event-log.md with date, variant, override category, decision, and rationale.

Varia is built by one person, Eric Gebauer, who is not a clinical geneticist or genetic counselor. Quality and rigor come from published criteria, measured residual human steps, and a transparent override surface, not from hidden solo judgment on each variant. Varia is not a substitute for a consultation with a clinician.

The named institutional sources Varia consults appear on the Institutional Authority References page with the date Varia last reviewed each. No institution has reviewed, endorsed, or partnered with Varia.

If you have questions about a specific finding in your scan, discuss it with your clinician. The Variant Evidence Summary PDF generated from your scan is designed to support that conversation with cited sources and facets, not replace clinical judgment.

Every variant in Varia goes through a published grading process that runs by software under human-authored rules, not by hidden individual review. See the sections above on human-in-the-loop grading, measurement integrity, the public scoring trace, and the citation library for how that works day to day.

The published rules go first, but they do not replace human judgment in every case. When the rules cannot make a confident decision, the case surfaces for human review per documented triggers. Every human override is logged publicly with measured counts.

Varia is built by one person, Eric Gebauer, who is not a clinical geneticist or genetic counselor. Quality and consistency come from published criteria, quantified residual human steps, and an auditable override log, not from one person's undocumented case-by-case calls. Varia is not a substitute for working with a clinician.

The institutional sources Varia consults appear on the Institutional Authority References page with the date each was last reviewed. No institution has reviewed, endorsed, or partnered with Varia.

If you have questions about a specific finding in your scan, discuss it with your clinician. The Variant Evidence Summary PDF generated from your scan is designed to support that conversation with cited sources and facets, not replace clinical judgment.