What it gives you
- Local models load on demand and retire after idle grace; first use and later reloads take longer than warm reuse
- Answers come from your documents, not the open web
- Citations link to the exact source paragraph
- Verification binds each claim to an evidence ID, line range, and content hash
- Supported, contradicted, insufficient, and uncertain verdicts stay explicit
- Below full substantive-claim support, the draft is withheld and the gap is named
- Verifier unavailability and capability degradation are reported, never silently downgraded
- Across 22 paired fixture tasks, verification lost no answer accuracy and produced strictly fewer unsupported claims
- Built-in local models need no cloud API or API key
- Optional HTTP inference offloads specific model roles while the corpus stays local
- Per-collection model overrides for prose versus code