Documentation

Install a verified release, run a first scan, sign and verify the report, or build from source

Take a model through the airlock on an intake host, with every command, and the console

The intake host for an on-prem GPU cluster: its shape, the one-time setup, bringing a model in, reading the clean store, and the record

Checksums, build provenance, the SBOM, offline verification, and a rebuild from source

Pull, ingest, promote, the trust policy, the log, the signed inventory export

The Helm chart: the airlock and the wizard in a cluster, who can reach them, and handing promoted models to LLMKube

The local web app over the engine: the scan wizard and the airlock console

The local HTTP API the wizard and the console call

Serving the models the store admits, read-only, over the Hugging Face Hub API, so an inference server downloads from the clean store

Rebuilding a pickle checkpoint as safetensors without unpickling it, and the signed derivation record

Attach and verify, the binding to the manifest digest, which layouts are matched, cosign and Kyverno

The Kyverno policies: which pod fields they read, what they check and what they do not, KServe, image volumes, and the policy-controller gap

The inputs that move each row: mirrors, denylists, provenance, signatures, a license policy, acceptances

Your organization's rules, which can only withhold: allowed formats, publisher allow and deny lists, required publisher signatures

Your organization's own known-bad hashes, reviewed chat templates, and canonical tokenizers, signed for air-gapped import: the format, verification, and building one

Fingerprint every tensor, and compare a model's weights with the reviewed model's, tensor by tensor

What Socair reads in an ONNX model: tensors, external data, operator domains, and metadata

What Socair reads in an OpenVINO IR model: the graph, the ranges it reads in its .bin, and rt_info text

What Socair reads in a Keras model: the HDF5 layout, a .keras archive's members, and code in the config

Opt-in, and no real Tier 2 check ships yet: the node-class record, the probe-helper protocol, isolation, and the roadmap

Which rows and features are evidence for which framework controls (MITRE ATLAS, OWASP LLM, NIST, AI supply-chain guidance), full or partial, and why

What each field holds and when it is left out, the G7 SBOM for AI minimum elements, and signing a BOM

The log for code-scanning tools: what each result holds, how a finding's fingerprint is computed, and uploading to GitHub code scanning

The attestation, section by section: the source of truth for what a report contains

The version every report records: what a version covers, when it changes, and the rule changes in each one

What Socair protects, whom it trusts for what, and what it does not defend against

Detection on reproductions of published attacks, misses pinned, beside picklescan, ModelScan, ModelAudit, and Fickling

How each detector was measured against real models