Skip to content

Latest commit

 

History

52 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Open Model Integration Validator (OMIV)

OMIV is an offline-first framework for verifying AI model artifacts, transformations, deployments, and runtime identity.

CI Python 3.11+ License: Apache-2.0 Status: public preview

Know what model artifact you have, what changed, what evidence supports it, and what remains unknown.

It is for model release teams, inference/runtime maintainers, conversion and quantization vendors, chip and platform teams, and enterprise AI teams.

It verifies bounded evidence for artifacts, shards, GGUF comparisons, transformations, configuration, trust, policy, and runtime records.

Try the offline quickstart below, then see the real Unsloth Gemma 4 case study.

OMIV is a pre-1.0 Alpha public preview, not a certification authority, safety evaluator, authenticity oracle, or production compliance product.

Why OMIV

A model name is not a model identity. A signature does not prove which weights a runtime loaded or produced an output.

OMIV separates declarations, observations, cryptographic verification, policy decisions, and unavailable evidence. Missing evidence stays unavailable.

Canonical validity is not authenticity. Structural parity is not behavioral equivalence. OMIV does not diagnose or fix every integration problem.

What OMIV verifies

Evidence surface Verifiable within scope Not established
Artifacts Bytes, manifests, structure, digests Safety or publisher authority
Shards/GGUF Completeness and structural comparison Universal compatibility
Changes Lineage, mappings, representation, samples Complete semantic equivalence
Trust Signatures, authority, custody, policy Policy appropriateness
Runtime Supplied resolution, deployment, output bindings Active weights without evidence

An absent, unsupported, or unchecked item never becomes PASS.

30-second quickstart

From a clean clone, install from source. Installation may contact a package index; the verification command itself is offline.

python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
omiv --version
omiv runtime-resolution verify runtime-resolution-parity/scenarios/immutable-pinned.json

Expected output and exit status:

0.10.0
VALID_CANONICAL_RUNTIME_RESOLUTION_OBJECT schema=omiv.runtime-resolution-scenario-result.v1
exit code: 0

The input is synthetic. The command verifies parsing, schema support, and identity/digest relationships for that object. It makes no provider request, resolves no live alias, observes no deployment or loaded weights, and runs no inference.

See the complete quickstart and offline demonstration.

Evidence chain

flowchart TD
    A[Source or registry] --> B[Acquisition evidence]
    B --> C[Artifact identity]
    C --> D[Manifest and shards]
    D --> E[Transformation evidence]
    E --> F[Signature and policy]
    F --> G[Deployment binding]
    G --> H[Output provenance]
    A -. declaration is not observation .-> B
    C -. artifact is not runtime identity .-> G
    E -. finite evidence is bounded .-> H
    B -. unavailable is not PASS .-> H
Loading

This is possible evidence flow, not universal automation. See architecture.

Current capabilities

OMIV currently covers:

  • artifact identity and manifests;
  • shard completeness and reconciliation;
  • GGUF/source-to-target comparison;
  • transformation and quantization evidence;
  • tokenizer/configuration evidence;
  • signatures, trust, and policy;
  • deployment, runtime, and output provenance.

Released boundaries remain narrow:

Phase Released scope Boundary
Phase 5 Passports, custody, attestations, trust, governance, security, runtime continuity, historical audit Supplied evidence; no automatic approval
Phase 6A Local manifests and byte identity No publisher or semantic claim
Phase 6B Pinned metadata and shard reconciliation Collection stays separate
Phase 6C Representation and bounded numerical-fidelity evidence Declared samples and limits only
Phase 6D Tokenizer/configuration parity and supplied probes Finite scope only
Phase 6E Runtime resolution, deployment binding, supplied results, output provenance No implicit live inference

Phase 6F Assurance Bundle interoperability is planned, not implemented. See the roadmap.

Reading results

Keep subject, scope, source, policy, and limits together. Unavailable stays unavailable. Signatures do not prove authority or runtime use.

Offline and fail-closed posture

Verification defaults to offline operation. Collection is explicit and separate; verification does not contact providers to fill gaps.

Missing evidence, unsupported schemas, integrity errors, and policy failures remain distinct.

The offline walkthrough covers tracked Phase 5 and Phase 6A–6E records and exit-code semantics.

Collectors and conversion runners may execute tools or access networks. They are outside the quickstart.

Real-world case study

Case Study 01 — Unsloth Gemma 4 E2B IT Q8_0 GGUF is an independent, offline structural and provenance-observability analysis.

The export succeeded. Current main-GGUF and mmproj sizes/SHA-256 values match retained historical C1 observations. Source-revision and format-native companion binding remained unavailable.

The historical C1 bytes are unavailable, so this is HISTORICAL_C1_MATCH, not a new direct C1/C2 file comparison.

The study makes no semantic-fidelity, numerical-fidelity, or runtime-compatibility claim. Structural validation is not behavioral equivalence.

We warmly thank Daniel Han for suggesting Gemma 4 E2B IT and the opportunity to test a real Unsloth export. OMIV worked independently; the exchange does not imply partnership, endorsement, approval, certification, or joint work.

Read the case study with its claims, evidence, results, and limits.

Practice-profile limitations

Practice profiles are bounded examples, not partnerships, endorsements, or claims about current provider state. See the technical reference.

Public and future commercial boundary

The public open core owns schemas, canonicalization, evidence semantics, signatures, offline verification, portable evidence, policy results, and public Assurance Bundle verification.

Possible future commercial work may add managed collection, private registries, monitoring, organization policy, RBAC/SSO, KMS/HSM, deployment admission, connectors, history, and support.

No commercial product is claimed. Future services may not redefine public OMIV semantics. See the full boundary.

Documentation

Start with the documentation index.

Detailed command and phase guidance lives in the documentation.

Project status and roadmap

The repository is public. OMIV remains a pre-1.0 Alpha public preview, not a production, certification, safety, or authenticity claim.

Phases 5 and 6A–6E are released in history. Phase 6F is planned, not implemented. The signed annotated v0.10.0 tag and GitHub pre-release exist; the PyPI project/version remain absent after a pre-publish failure.

Public availability is not release completion. See the roadmap and release notes.

Contributing, security, support, and license

Contributions follow CONTRIBUTING.md and GOVERNANCE.md.

Report vulnerabilities through SECURITY.md. Use SUPPORT.md for the support boundary.

OMIV uses Apache-2.0. See NOTICE, third-party notices, and trademarks.

OMIV is independent and is not an official, approved, endorsed, affiliated, or certified tool of any model provider.

About

Offline-first evidence and verification framework for AI model artifacts, transformations, runtime identity, and provenance.

Topics

Resources

Contributing

Security policy

Stars

2 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages