Guardrails — why it can't hallucinate its way into trouble
Schema-first: outputs validated + repaired against a declared JSON schema
Confidence gate: low-confidence → deterministic floor, never a guess
Grounded: answers constrained to signed pack knowledge, not open web
HITL consent: side-effecting tools (payments, messages) held for approval
Glass-box trace: every route, skip, tool call and token logged with an id
Eval gate: packs must pass an eval suite before signing/publish
Run a task to see the guardrail trace…
Central nervous system — three tiers, one mesh
this device · Tier 1 Edge (TLM/OS-AI)──▶
vehicle-node-3 · Tier 2 Node (SLM + local RAG)──▶
@cloud.hub · Tier 3 Core (LLM)
field-tablet-02
kiosk-07
Tier 1 handhelds structure inputs locally (no raw audio upstream) → Tier 2 vehicle/site nodes aggregate, dedupe and answer from regional RAG → Tier 3 core consolidates high-signal summaries into business memory (hierarchical RAG, store-and-forward sync). Heavy compute only when deep reasoning is required — and the hub can never bypass on-device consent.