Core Concepts
Understand the building blocks of VeldrixAI: Trust Scores, Pillars, Policies, and Enforcement Actions.
Trust Score
Every evaluation produces a Trust Score — a float between 0 and 1. A score of 1.0 means the response is clean across all evaluated pillars. A score below your policy threshold triggers an enforcement action.
Pillars
VeldrixAI evaluates responses through five independent pillars. Each pillar runs in parallel — total latency equals the slowest single pillar, not the sum of all five.
Policies
A policy binds together: which pillars to run, what thresholds to use, and what enforcement action to take on violation. Policies are versioned and can be deployed without downtime.
Enforcement Actions
Seven enforcement actions are available, listed by priority:
| Action | Priority | Effect |
|---|---|---|
escalate | 1 | Hold response; route to human review queue |
block | 2 | Suppress response; return configured fallback |
rewrite | 3 | Replace with policy-safe version |
regenerate | 4 | Re-invoke LLM with adjusted prompt |
mask | 5 | Redact PII/sensitive entities in-place |
disclaimer | 6 | Append legal or regulatory disclosure |
allow | 7 | Pass response through unchanged |
Evaluation Modes
Inline mode: Evaluation happens synchronously before the response is delivered. Adds 10–50ms P95 latency.
Background mode: Response is delivered immediately; evaluation runs async. Adds zero latency to your critical path. Violations trigger post-delivery enforcement (e.g., audit log entry, escalation).