QAVE — Sovereign AI Assurance

QAVE (Qurator Assurance Vertical Engine) separates generative reasoning from deterministic reward scoring:

1. Reasoning lane — criteria-aware critic extracts grounded evidence (evidence_v0) with optional Knowledge Base, annotations, and ground_truth.
2. Scoring lane — calibrated 4B / Lightning 1.7B reward heads produce a stable score in [0, 1] against operating-point τ*.
3. Context — up to 350K OneGulp entitled windows on Assurance Deep paths.
4. Deployment — Freemium / Paid self-serve or air-gapped Sovereign (unmetered contract keys).

Core surfaces

Studio loop

Setup → Assure → Calibration → Insights — see Getting Started.

Continue with the Criteria Writing Guide or jump to a cookbook.