# Advisory AI overview

Advisory AI provides explainable summaries and prioritization hints for human
review. It consumes canonical observations from Concelier and Excititor and
returns evidence-linked outputs.

Inputs
- Advisory observations and linksets from Concelier.
- VEX observations and trust metadata from Excititor.
- Optional SBOM context snapshots for component relevance.

Outputs
- Human-readable summaries with evidence references.
- Priority hints and rationale tied to source ids and hashes.
- Payloads suitable for UI and CLI display, not for policy decisions.

Guardrails
- No uncontrolled external network calls.
- Redaction of sensitive fields before model input.
- Outputs must include source ids and evidence hashes.
- Deterministic templates for offline or restricted modes.

Evidence payloads
- Each response includes model metadata, input hashes, and output hashes.
- Evidence links are preserved so decisions remain auditable.

Packaging and offline
- Offline kits bundle prompt templates and model policies.
- Model use is gated by configuration and approval policies.

Related references
- docs/advisory-ai/overview.md
- docs/advisory-ai/guardrails-and-evidence.md
- docs/advisory-ai/evidence-payloads.md
- docs/advisory-ai/sbom-context-hand-off.md