SovAIHub
ModulesSAI-200
SAI-200 table of contents
Concept2 min readContent reviewed

Internal repositories and model stores

Define authoritative internal sources for containers, packages, models, metadata, and dependency resolution.

Last content review 2026-08-11Included in SAI-200

Before you start

Module prerequisites: SAI-110, SAI-120. Check the exercise-specific tools and environment before running commands.

Module outcomes: Design an approved import workflow; Use checksums, signing, SBOMs, registries, and wheelhouses; Plan repeatable disconnected builds and updates. This lesson contributes to that sequence.

Internal repositories become authoritative

Disconnected builds must not search uncontrolled sources at build or run time. Define internal repositories for each artifact class: container images, operating-system packages, language packages, models and tokenizers, datasets, charts or manifests, policies, and configuration. A repository is authoritative when it is the only source a build or deployment is permitted to resolve from — not merely a convenient mirror alongside the public internet.

Repository design

Document:

  • Namespace and naming rules, immutable identity, metadata, and lifecycle state.
  • Which sources may publish, promote, consume, deprecate, or delete, and under whose authority.
  • How dependencies are closed over, and how a missing dependency fails — silently substituting an external source is not an acceptable failure mode.
  • Isolation between quarantine, staging, and approved channels, so an unreviewed artifact cannot be resolved by a production build.
  • Retention, replication, backup, restore, capacity, and integrity verification.
  • Revocation, and how consumers are prevented from selecting revoked content once it is marked unsafe.
  • Export and portability needed for recovery or a future platform change — see Portability, reversibility, and exit.

Prefer immutable versions and digest pinning over mutable tags. A mutable tag may still be useful as a discovery aid, but it must resolve to the identity recorded by the release, and that resolution must be logged.

Models need more than object storage

A model store should connect weights and tokenizer to license, provenance, evaluation, runtime compatibility, serving configuration, approval scope, and known limitations. Storing a model file without this decision context turns later use into guesswork — a future consumer has no way to know whether the file in front of them was ever actually approved for the purpose they intend.

Test repository unavailability, corruption, restore, expiry, and revocation directly, not only in design review. A local mirror is not authoritative until its governance and operational ownership are explicit and its failure behavior has been exercised.

Module worked example

A model archive arrives without a source revision or approved digest.

Apply the control
Keep it in quarantine, obtain provenance, inspect the package and record approval before promotion.
Expected result
The promoted artifact has a traceable identity, review record and rollback location.
Troubleshooting
A familiar file name is insufficient. Compare the digest and manifest against the approved record.

Practice and completion check

Choose one control from this lesson. Identify its actor, target, enforcement point and expected result. Describe an allowed case and a denied or failed case. For a runnable exercise, retain observed output and separate it from expected output. Record unresolved limitations before marking the lesson complete.

Your lesson progress

Completion is self-reported learning progress; it does not certify practical competence.

Status: Not marked

My learning

Loading your learning record…