Skip to content

Runtime Check & Correction

7. Runtime Check & Correction

The Runtime Check & Correction pipeline is the execution engine that enforces Guard rules in real time. It intercepts, evaluates, and corrects AI interactions inline.

7.1 Three-Stage Enforcement Pipeline

Enforcement is processed as a sequential three-stage pipeline:

  1. Precheck: Rapid initial evaluation of the incoming prompt before it reaches the target AI model: runs Zero-Data AI Data Plane (ZT-ADP) format-preserving encryption and blinding for sensitive data, and performs L1 Precedent Decision Cache matching for sub-10ms fast-deny of known non-compliant inputs.
  2. Check: Evaluation of the AI model's response in parallel with the generation stream to detect compliance violations via L2 Sentinel Fast and, if escalated, L3 Sentinel Max/Ultra or G1 MoE.
  3. Correction: Inline rewriting or modification of the response to align with policy constraints before delivery to the end user.

7.2 <200ms SLA

The sync path is highly optimized to run under a strict sub-200ms sync path SLA to ensure that real-time enforcement does not degrade the user experience or introduce noticeable latency.

7.3 Enforcement Actions

When a Guard triggers, the system applies one of four actions at inference time:

  • Allow: The interaction complies with policy and is delivered normally.
  • Rewrite: The output is corrected or sanitized inline to remove compliance violations.
  • Escalate: The interaction is flagged and routed to the Route → Runtime Reviews queue (optionally using a Safe Degrade response in the interim).
  • Block: The interaction is stopped, and a policy violation notice is returned.