Local determines where AI runs. Locs determines whether it is safe, grounded and approved.
A local model can still hallucinate, leak identifiers, follow injected instructions, cite the wrong source or regress after an upgrade. The Assurance Lab treats every model, embedding model, retrieval configuration and prompt policy as a governed component that is evaluated, approved, activated, monitored and rolled back.
Private deployment answers where the AI runs. It does not answer whether the AI is grounded in approved knowledge, whether it refuses dangerous requests, whether it leaks sensitive identifiers, or whether it still behaves after a model or knowledge update. Those are measured, not assumed.
Versioned, checksummed
Evaluation datasets move through a governed lifecycle (draft → review → approved → active) and are checksummed for tamper evidence.
Grounding & citations
Measures whether answers are grounded in the right sources and whether required citations are present.
Refusal recall
Measures whether the system refuses when approved knowledge is insufficient or the request is dangerous.
Sensitive-data tests
Probes for leakage of sensitive identifiers under adversarial prompts.
Prompt-injection tests
Measures resistance to injected instructions embedded in retrieved content. We measure it; we do not claim perfect protection.
Staleness & rollback
Flags configurations that drift, and every activation creates a rollback relationship so you can revert.
A configuration is not trusted until it is proven
An Assured Configuration binds the model, embedding model, retrieval configuration, prompt policy, knowledge-pack versions and the golden set it is evaluated against into one immutable, deployable unit. Its lifecycle is governed:
draft → queued → running → completed → needs_security_review → approved → active → suspended / rolled_back
A failed critical threshold blocks approval. The creator cannot be the sole approver (two-person rule). Only one active configuration per module and deployment profile. Activation and rollback emit signed execution receipts, and a kill switch blocks serving regardless of assurance state.
See the assurance methodology applied to your knowledge.
The Lighthouse Pilot authors a customer-specific golden set and runs a full assurance evaluation before any operational use.