# Data sources and provenance

## Planned authoritative sources

| Source | Role | Access strategy |
|---|---|---|
| OMIA | Canine disease-gene and causal-variant evidence | Official SQL/XML dump and NCBI gene mapping |
| HPO | Human phenotype ontology and annotations | Pinned release files |
| HGNC | Stable human identifier crosswalk | Checksum-pinned complete-set export |
| Monarch / PHENIO | Cross-species mappings and disease evidence | Dated Monarch subsets plus a compact pinned PHENIO closure |
| MGI | Mouse gene-phenotype and homology evidence | Official reports |
| Ensembl | Dog-human-mouse orthology | Pinned bulk export or cached REST ingest |
| Dog10K | Population and breed-aware frequency evidence | Later compact derived index from official BCF files |

## Required provenance fields

Every processed row must remain traceable through:

- source identifier;
- source record identifier;
- source release or retrieval date;
- source URL;
- citation;
- evidence species;
- transformation version;
- genome assembly for coordinate-bearing records.

## Assembly policy

Coordinates from `UU_Cfam_GSD_1.0`, CanFam3.1, ROS_Cfam, or another assembly are never mixed silently. Liftover results must retain both original and target coordinates plus the chain-file version and mapping status.

## Gene identity and orthology policy

Ensembl and NCBI Gene identifiers are the stable join keys. Gene symbols and aliases are retained for display and search, never used alone to merge records. Dog-human and dog-mouse edges retain their Ensembl release, species annotation assemblies, homology type, confidence, ambiguity status, and source record ID.

The checked-in lightweight orthology manifest remains in an `unmaterialized` state so ordinary source checkouts cannot accidentally claim production evidence. The Ensembl 116 release descriptor pins three official genome-specific exports, their MD5 and SHA-256 digests, byte counts, assemblies, and exact expected normalized counts. `pipelines/ingest_ensembl_orthology.py` verifies and unions those partitions into a `pinned` snapshot. See [Gene identity and orthology](ORTHOLOGY.md).

## HPO release policy

The HPO module pins ontology release `v2026-09-01` and the matching `2026-09-02` disease annotations by immutable GitHub release asset ID, byte count, and SHA-256. It materializes ontology, gene-phenotype, and disease-phenotype JSONL files plus a checksum manifest. The snapshot is explicitly `current-snapshot-only`; it is ineligible for temporal evaluation. See [Human Phenotype Ontology ingestion](HPO.md).

## Monarch release policy

The Monarch module uses the dated `2026-09-02` KG association subsets, not the moving `latest` path. A separately pinned HGNC export resolves human HGNC identifiers to Ensembl and NCBI Gene IDs without symbol joins. Deterministic evidence IDs support pre-feature masking, and exact duplicate source rows are collapsed. Integrated MGI rows retain their primary-source attribution and cannot be counted independently from direct MGI evidence. See [Monarch Knowledge Graph ingestion](MONARCH.md).

## MGI release policy

The direct MGI module checksum-pins four moving official reports as a retrieval-dated snapshot and verifies the embedded MP ontology release. MGI-to-NCBI mappings from two reports are reconciled explicitly: disagreements and unresolved markers remain visible but are not rank-eligible. Direct MGI and Monarch's MGI-derived associations are provenance-linked inputs, not independent evidence. See [Direct MGI ingestion](MGI.md).

## OMIA release policy

The OMIA module checksum-pins the moving MySQL dump and its separate official internal-to-NCBI Gene export as a retrieval-dated pair. It materializes only the canine subset, never joins on symbols, retains unresolved gene mappings for audit, and excludes those rows from ranking. Deterministic evidence and binomial OMIA record IDs are available for pre-feature discovery masking. The snapshot is current-only and cannot support temporal evaluation. See [Canine OMIA ingestion](OMIA.md).

## PHENIO release policy

The PHENIO module pins release `v2026-09-01`, its source commit, immutable relation-graph and upstream-version asset IDs, byte counts, and SHA-256 digests. It scans the complete entailed relation graph but emits only the reflexive HP/MP/uPheno subclass closure required for deterministic cross-species similarity. The upstream ontology version table is retained intact. This current snapshot cannot support historical temporal evaluation. See [PHENIO / uPheno ingestion](PHENIO.md).

## Hugging Face publication

`pipelines/build_hf_release.py` creates a deterministic release beneath `dist/` from an explicit allowlist. The output includes viewer-friendly benchmark JSONL, the selected validated orthology snapshot, optional verified HPO, Monarch, MGI, processed canine OMIA, and compact PHENIO snapshots, provenance manifests, schemas, documentation, and a checksum manifest. `pipelines/publish_huggingface.py` re-verifies the package and uploads it only to a public Hugging Face Dataset repository named explicitly by the operator.

Publication is manual. `HF_TOKEN` is read only from the environment or a GitHub Actions secret; it is never accepted as a command-line value or stored in a generated file. Complete upstream database exports and source-link-only artifacts are excluded.

## Demonstration snapshot

The repository currently ships a deliberately small hand-curated ontology and evidence snapshot. `VH:*` identifiers are internal demonstration terms, not a replacement veterinary phenotype ontology. Production data will preserve HPO, MP, MONDO, OMIA, NCBI Gene, and Ensembl identifiers.