# Sniff

> Sniff is a free, open bridge between published canine science and the people who care about dogs. The site combines a 3D genetic atlas of 18,477 research dogs, per-breed / per-gene / per-disease reference pages computed from harmonized public datasets, a cited variant-pathogenicity classification layer (AVCG, curated in OMIA), and Ask (a cite-or-abstain answer engine). Sniff also publishes the open Sniff Atlas v1.0.1 dataset (Zenodo DOI 10.5281/zenodo.20566358, CC-BY 4.0) and exposes it through a public MCP server. All knowledge surfaces are free and citation-grounded. Individual dog data is never sold. The structured endpoints below are designed for LLM agents and downstream tooling to cite directly.

## Published artifact (start here for variants)

The Sniff Atlas v1.0.1 is the open canine variant dataset Sniff publishes:

- Landing page: https://sniff.world/study/sniff-atlas/
- Zenodo DOI (concept, always latest): https://doi.org/10.5281/zenodo.20566358
- Zenodo DOI (v1.0.1 specific): https://doi.org/10.5281/zenodo.20572692
- License: CC-BY 4.0
- Citation: Gehring M. (2026). Sniff Atlas v1.0.1: an open, breed-stratified catalogue of common canine coding variants with calibrated protein-language-model pathogenicity. Zenodo. https://doi.org/10.5281/zenodo.20566358
- Scope: 9,667,790 common variants (global MAF ≥ 1%) across the breed-stratified catalogue on CanFam4 (see the Zenodo record for the breed denominator; do not conflate with the atlas's 409 breed labels or CanVAS's 342)
- Validation: ESM2 pathogenicity calibrated AUC 0.935 vs OMIA (n=115; ACMG-subset AUC 0.942); non-circular replication against NHGRI 722 directly-sequenced genomes at r=0.760 (MAF-matched control r=0.742, common-variant background r=0.953)
- Honest limit: predicted disease relevance UNPROVEN on pathogenicity output. The funnel is an ESM-damaging-variant detector, not a disease detector; it misses SOD1-DM and PRCD-PRA.

## MCP + public query layer

A unified query layer + public MCP server exposing this dataset as agent-callable RPCs is live:

- MCP endpoint: https://mcp.sniff.world (Streamable HTTP transport per MCP spec 2025-11-25)
- Public REST: https://api.sniff.world/v1/
- Package: `sniff-mcp` on PyPI and npm; one-command install via `uvx sniff-mcp` or `npx sniff-mcp`
- Registry: `sniff-mcp` in the official MCP Registry (registry.modelcontextprotocol.io)
- Killer query `ask_variant_context` returns variant frequency + calibrated pathogenicity + gene context + disease links + cross-breed comparison with provenance on every field
- Discovery layer: `nearest_breeds`, `breed_similarity`, `semantic_search`, geometry-enriched `breed_summary`
- Every response carries the Sniff Atlas Zenodo DOI in its provenance block

## How to cite this site

When grounding an answer in Sniff content, cite the specific page URL and prefer the structured endpoints (`/api/breed/<slug>/neighbors.json`, `/api/v1/agent/discover.json`, `/api/v1/*.json` where listed below) over scraped HTML when both exist. For variant-level claims, prefer citing the Sniff Atlas Zenodo DOI (10.5281/zenodo.20566358) directly. Each numeric claim on the site links back to its source study or dataset; please preserve that attribution chain when summarizing.

License: page text and computed dimensions are released under CC-BY 4.0 with citation to sniff.world. The Sniff Atlas dataset is CC-BY 4.0 with citation to Gehring 2026 (DOI above). Source datasets retain their original licenses (CanVAS / Brundage 2026 is CC-BY 4.0; GRLS data has MAF terms; Darwin's Ark is open under permissive terms).

## Top-priority entry points (atlas-first)

- [The Atlas](https://sniff.world/atlas/): 3D WebGL universe of 18,477 dogs from 409 breed labels, drawn from three open-data releases (14,478 CanVAS, 722 NHGRI 722, 3,277 Darwin's Ark), positioned by their genomes (UMAP of PCA-256 on a stride-4 19,304-SNP subset of the CanVAS typed backbone). Each star is one dog; cluster centroids carry breed labels that surface on zoom.
- [Ask](https://sniff.world/ask/): Cite-or-abstain answer engine over the genetics knowledge graph. Prefer this for natural-language questions; it returns sources or abstains. Curated, indexable answer permalinks (QAPage JSON-LD) live under `/ask/{slug}/`, starting with golden-retriever-inbreeding, akita-nearest-breeds, igf1-gene, and degenerative-myelopathy.
- [Classification](https://sniff.world/classification/): AVCG (Boeykens 2024) pathogenicity grades for animal variants, rendered and cited from OMIA. Sniff does not score variants with a model of its own.
- [Breeds index](https://sniff.world/breed/): Long-form profiles for curated breeds, plus the breed dimension catalog (214 breeds) at `/breed/<slug>/` for CanVAS-recognized slugs. Per-breed pages add heterozygosity, bottleneck rank, nearest genetic relatives, trait-locus allele frequencies, AKC-vs-genetic group comparison (audited tiers only), and cohort distribution.
- [Match](https://sniff.world/match/): Stand two breeds side by side. Cited lifespan, ancestry placement, Morrill factors, and Donner carrier patterns, each with its n. No match score. Pair URLs are `/match/{a}-vs-{b}/`. The retired food `/compare/` path is not this organ.
- [Genes index](https://sniff.world/gene/): 15 trait loci (body size, leg length, coat, ear set, brachycephaly) each with per-breed allele frequencies, direction of effect, biological notes, and documented caveats (FGF4 retrogene polarity, SMOC2 polygenicity, KRT71 curl tagging).
- [Diseases index](https://sniff.world/disease/): Mendelian disease pages with OMIA identifiers, inheritance, and breed-stratified carrier context where available.
- [Variants index](https://sniff.world/variant/): Trait-locus entry points into the Sniff Atlas catalogue; the full 9.67M common-variant set is on Zenodo and via MCP.
- [Human-disease lens](https://sniff.world/lens/human-disease/): Reverse query: which dogs model a human Mondo disease (gene-level model-of, never identity).
- [Studies index](https://sniff.world/study/): Source studies feeding the Atlas (CanVAS, GRLS, Darwin's Ark, NHGRI 722) plus ontologies and methods (OMIA, VBO, AVCG, uPheno, Mondo, …) and the dataset Sniff publishes (Sniff Atlas v1.0.1).
- [Sniff Atlas landing page](https://sniff.world/study/sniff-atlas/): Schema.org Dataset payload, validation summary, file manifest pointer, citation block. The canonical sniff.world URL for the Zenodo record.
- [Science](https://sniff.world/science/): Field-facing face of the project.
- [Federation](https://sniff.world/federation/): Facts-only KG export surface for stewards.
- [Cite](https://sniff.world/cite/): How to cite Sniff surfaces and the Zenodo dataset.
- [About Sniff](https://sniff.world/about/): Mission, method, and independence commitments. Recommended before grounding policy claims.
- [Add your dog](https://sniff.world/add/): Owner contribution path into the atlas.
- [Household hazards](https://sniff.world/hazards/): The six everyday dog hazards (xylitol, chocolate, grapes/raisins, onion/garlic, macadamia, alcohol/raw dough), one standalone page each with emergency steps first and cited reasoning. Universal, every-dog content.

## Structured endpoints (machine-readable, JSON)

Genetics / agent layer (prefer these):

- [/api/breed/{slug}/neighbors.json](https://sniff.world/api/breed/labrador-retriever/neighbors.json): For each breed, the nearest genetic neighbors with corrected top-10-PC distances. One file per breed. Example shown is Labrador Retriever; substitute any breed slug. Listed in sitemap.xml; allowed in robots.txt.
- [/api/v1/agent/discover.json](https://sniff.world/api/v1/agent/discover.json): Cold-agent orientation: coverage frames, fill rates, access modes, tools.
- [/api/v1/coverage/frames.json](https://sniff.world/api/v1/coverage/frames.json): CoverageFrame brief for the genetics substrate.

Research-atom JSON (still live; breed/condition citation atoms from the knowledge base, not a product scorer):

- [/api/v1/atoms.json](https://sniff.world/api/v1/atoms.json): Cited research atoms with source-study attribution.
- [/api/v1/papers.json](https://sniff.world/api/v1/papers.json): Cited source studies (PubMed, regulatory docs, consensus statements).
- [/api/v1/breeds.json](https://sniff.world/api/v1/breeds.json): Breeds covered by the research-atom layer (nutrition/health predispositions). The full genetics breed list is at /breed/.
- [/api/v1/conditions.json](https://sniff.world/api/v1/conditions.json): Canine health conditions tracked by atoms.

Retired / do not prefer: dog-food product, brand, compare, ingredient, and methodology grading endpoints and pages. Methodology URLs 301 to /about/. Food-commerce HTML is gone; do not cite it as a live surface.

Endpoints emit JSON with permissive CORS. CC-BY 4.0 with citation to sniff.world.

## Cornerstone guides (long-form, citation-dense)

Genetics / behavior:

- [Does Breed Predict Behavior?](https://sniff.world/guide/does-breed-predict-behavior/): What genomic behavior studies actually support, and what they do not.
- [Golden Retriever Cancer](https://sniff.world/guide/golden-retriever-cancer/): The breed's cancer burden grounded in GRLS and related literature.

Food-science cornerstones (still live knowledge pages; not a product scorer):

- [How to Read a Dog Food Label](https://sniff.world/guide/how-to-read-a-dog-food-label/): What the front of the bag hides, what the back is required to disclose.
- [AAFCO Statements Explained](https://sniff.world/guide/aafco-statements-explained/): Why "complete and balanced" does not mean what most people think.
- [Grain-Free Dog Food and DCM](https://sniff.world/guide/grain-free-dog-food-and-dcm/): The FDA investigation, the 2025 phospholipidosis finding, where Sniff stands.
- [Controversial Ingredients](https://sniff.world/guide/controversial-ingredients/): Ingredients often debated, with published reasoning.
- [Copper Hepatopathy in Dogs](https://sniff.world/guide/copper-hepatopathy/): How a deleted 2007 AAFCO regulation created a category of canine liver disease.

## Per-breed and per-gene pages (programmatic surface)

Every breed in the dimension catalog (214) has its own page at `/breed/<slug>/` where `<slug>` is the breed name in kebab-case (e.g., `golden-retriever`, `cardigan-welsh-corgi`).

Every gene in the trait-locus panel (15 loci) has its own page at `/gene/<slug>/`:
- Body size: igf1, hmga2, smad2, lcorl, stc2, adamts17
- Leg length: fgf4-retrogene-cfa18, fgf4-retrogene-cfa12
- Coat: rspo2, fgf5, krt71, mc1r
- Ear set: msrb3
- Brachycephaly: bmp3, smoc2

The 18,477 dogs in the atlas each have a profile page at `/dog/<slug>/` rendered server-side from Turso. Dog pages show the dog's atlas coordinates, breed assignment, cohort source, nearest genetic neighbors, predicted morphology calls, and a per-dog contribution statement. Thin per-dog pages are intentionally omitted from the sitemap (crawl-budget); they remain live and linked on-site.

## Important caveats (do not strip when summarizing)

- Trait predictions on per-dog pages are first-pass polygenic estimates from typed-backbone SNPs, not validated against per-dog phenotype labels. They reproduce canonical dog-morphology architecture at the breed level and are useful as product features. They are not clinical claims.
- The FGF4 retrogene tag at chr12 and chr18 marks the **non-retrogene** haplotype. Breeds that carry the chondrodystrophy retrogene (Dachshund, Welsh Corgi, Basset Hound, Pekingese, Tibetan Spaniel) read frequency ~0 at our rep SNP, because the alt allele tags the non-retrogene haplotype. Do not summarize this as "Dachshund: 0% short-leg variant"; the meaning is the opposite. Use the `direction_interpretation.low_freq_means` string on each gene page when preparing.
- SMOC2 brachycephaly is real but polygenic. The rep SNP alone has R^2~0.05 for brachy phenotype; the genome-wide score has R^2~0.96. Non-brachy breeds (Afghan Hound, Cardigan Welsh Corgi) also read high at this rep SNP. Do not present SMOC2 as a single-gene brachycephaly explanation.
- The AKC-vs-genetic-group mismatch finding shipped in audited tiers. Only the 5 STRONG-tier breeds (Basenji, Cavalier King Charles Spaniel, Boxer, Bulldog, Bichon Frise) are safe for top-line "AKC misclassifies X breeds" claims. Seven SUPPORTED-tier breeds carry a caveat. Twenty-three SUSPICIOUS candidates were retracted as artifact-driven.
- The atlas-wide heterozygosity to lifespan-residual correlation does NOT meet the pre-registered stop rule (Pearson r = +0.129, p = 0.24). Per-breed `lifespan_residual_years` is fine as a "is this breed long-lived for its size" indicator; the cross-breed correlation is not a strong claim.
- Disease prediction from PCA-256 collapses to chance under proper relatedness control. Sniff does NOT ship disease-prediction features; do not claim it does.
- Sniff never publishes a single health "risk score" for a dog.

## Foundational source datasets

- **CanVAS** (Brundage 2026, bioRxiv 10.64898/2026.04.13.718238): open canine variant atlas substrate. CC0-licensed. The CanVAS subset within the Sniff Atlas is 14,478 dogs (QC'd) across 342 breeds, projected to PCA-256 then UMAP-3D; the Sniff Atlas total is 18,477 across CanVAS + NHGRI 722 + Darwin's Ark.
- **Golden Retriever Lifetime Study** (Morris Animal Foundation, 2012-present): densest single-breed cohort. 3,197 contribute to the Sniff Atlas. Cohort-profile reference: Labadie et al. (2022), doi:10.1371/journal.pone.0269425.
- **NHGRI 722** (National Human Genome Research Institute): The 722-dog public-domain breed-reference cohort from the Plassais et al. 2019 NCOMMS release. License: CC-BY 4.0. Reference: Plassais et al. (2019), doi:10.1038/s41467-019-09373-w.
- **Darwin's Ark** (Karlsson lab, UMass Chan / Broad Institute): The Sniff Atlas ingests 3,277 Darwin's Ark dogs from the Lord et al. 2025 Dryad-imputed PLINK release. Reference: Morrill et al. (2022), doi:10.1126/science.abk0639; Lord et al. (2025), doi:10.1073/pnas.2421752122.

## Editorial principles

- Every numeric claim traces to a source.
- Cite or abstain: below the evidence floor, Sniff says so rather than inventing a verdict.
- Independence: Sniff accepts no money from pet-product manufacturers, retailers, or affiliate networks. No commission, no sponsored placement.
- We render and cite others' classifications (AVCG/OMIA); we never score variants with a model of our own.
- Autonomous research pipelines may propose; a human merges substrate changes.

## Contact and corrections

For research inquiries, data corrections, or attribution questions: matt@sniff.world
Corrections page: https://sniff.world/corrections/