Skip to content

Choose an interface

Start from what your integration does during an analysis.

You are adding… Implement Public import
live inference, a hosted model call, or a model-specific scoring plan ModelPlugin altar.models
precomputed per-variant evidence AnnotationSource altar.sources
one-to-many variant/regulatory-element-to-gene evidence VariantGeneLinkSource altar.sources
persistent canonical variant–gene evidence VariantGeneLinkStore altar.sources
container task execution on a compute system ExecutionBackend altar.execution
persistent model-score storage and result assembly ScoreStore altar.results
file/object movement Storage altar.execution
token authentication (provisional) AuthProvider altar.access
asynchronous job handoff (provisional) JobQueue altar.execution

AuthProvider and JobQueue are provisional: they are importable, but they lack a public production adapter and an Altar consumer, so their shape may change in a minor release.

Keep integrations composable

A scientific binding produces backend-neutral plans or evidence. An infrastructure implementation consumes a public interface. Do not create separate integrations for every model × compute × storage combination.

For example, a Cherimoya binding should work with local Docker, Modal, and Kubernetes through the same plan. An E2G importer should emit VariantGeneEvidenceRecord into VariantGeneLinkStore rather than becoming OpenTargetsPostgresSource, EncodeRE2GParquetSource, and similar source × database combinations with duplicated scientific normalization.

Model or source?

Ask whether inference occurs in this analysis:

  • Yes: use a model binding and declare how work runs.
  • No, values already exist: use an annotation or relation source and preserve the release identity.

The upstream origin of a dataset does not decide its Altar interface. Precomputed neural-network scores are still lookup evidence in an analysis that does not execute the network.

Definition of a supported extension

An interface implementation is ready when it has:

  • one real consumer or integration;
  • complete schemas and provenance behavior;
  • missing-data and error semantics;
  • entry-point registration;
  • the matching Altar conformance suite;
  • scientific and operational documentation;
  • isolation from private host-application imports.