Advisory AI Packaging & SBOM Bundle

This guide defines the artefacts, provenance, and deployment steps required to ship Advisory AI to on-prem and air-gapped (Offline Kit) installs. Every image, prompt pack, and policy bundle carries an SBOM and a DSSE attestation so operators can verify the supply chain before enabling inference.

Audience: release engineers assembling Advisory AI packages, and operators installing them into a tenant.

1) Bundle contents

ArtefactPurposeProvenance
advisory-ai-web imageAPI surface + plan cacheSBOM: SBOM-AIAI-31-001:web; DSSE attestation signed by Release key
advisory-ai-worker imageQueue + inference executorSBOM: SBOM-AIAI-31-001:worker; DSSE attestation
Prompt + guardrail packDeterministic prompts, redaction lists, validation rulesDSSE sealed; hash recorded in prompts.manifest
Profile catalogdefault, fips-local, gost-local, cloud-openai (disabled)Versioned JSON, hashed; tenant consent flags captured
Policy bundlepolicyVersion digest for baseline evaluation; Authority importableDSSE + provenance to Policy Engine digests
LNM evidence export (optional)Concelier advisory_linksets + advisory_observations for air-gap replayHash list aligned to provenanceHash in RAG bundles
SBOM context client configExample AddSbomContextHttpClient settings (BaseAddress, Endpoint, ApiKey)Signed sbom-context.example.json

2) Directory layout (Offline Kit)

/offline-kit/advisory-ai/
  images/
    advisory-ai-web.tar.zst
    advisory-ai-worker.tar.zst
  sboms/
    SBOM-AIAI-31-001-web.json
    SBOM-AIAI-31-001-worker.json
  provenance/
    advisory-ai-web.intoto.jsonl
    advisory-ai-worker.intoto.jsonl
    prompts.manifest.dsse
    profiles.catalog.json
    policy-bundle.intoto.jsonl
  config/
    advisoryai.appsettings.example.json
    sbom-context.example.json
  evidence/
    lnm-linksets.ndjson      # optional; aligns to linkset hashes in RAG bundles
    lnm-observations.ndjson  # optional; immutable raw docs

3) SBOM & provenance rules

4) Deployment checklist

5) Update obligations