Trust Engine Overview
The Trust Engine runs five evaluation pillars in parallel against every model output and returns a single, policy-weighted verdict.
Architecture
The Trust Engine is a stateless evaluation service. For each request it:
Tokenizes and preprocesses the prompt + response pair.
Dispatches all five pillars concurrently under a shared latency budget.
Aggregates pillar scores using your policy weights.
Resolves the verdict from your thresholds (plus any hard-block flags).
Returns the result and emits an audit record.
The pillars
Safety & Toxicity
Harmful, violent, harassing, self-harm content.
Hallucination
Fabricated facts and citations.
Bias & Fairness
Stereotyping and discriminatory framing.
Prompt Security
Injection and jailbreak detection.
Compliance & PII
Regulated data and personal information.
Scoring model
Each pillar emits a score in [0,1] and a confidence value. The overall trust score is the weighted aggregate of pillar scores. Because weighting is deterministic, the same input and policy version always produce the same verdict — essential for reproducible audits.
Latency tiers
You choose an SLA tier per request via the x-veldrix-sla-tier header (or your plan default). Per-pillar work is generous, but the total budget is the hard deadline:
| Tier | p95 budget | Use for |
|---|---|---|
REALTIME | ≤ 200 ms | Interactive, user-blocking calls. |
STANDARD | ≤ 500 ms | Default for most traffic. |
BACKGROUND | uncapped | Batch / fire-and-forget logging. |
Grounded evaluation
Pass source context (retrieved documents, system instructions) in the context field and pillars evaluate the response against that ground truth — sharply reducing hallucination false-positives versus context-free checking.