Skip to main content

ThirdLaw Intervene

ThirdLaw Intervene is the enforcement and response layer of the ThirdLaw platform. It translates violations of "Laws"—as defined and executed in the Evaluate service—into live enforcement actions. Intervene ensures that risks flagged by Evaluate are operationalized into real-time or batched mitigations, allowing IT and Security teams to exercise control over the behavior of LLM-connected systems.

Key Capabilities​

ThirdLaw Intervene provides mechanisms to respond to detected issues, from simple logging to response blocking or human review triggers. It offers flexible intervention options based on the severity and nature of detected violations.

  • Logging and Alerting: Recording and notifying about detected issues
  • Response Blocking: Preventing problematic responses from reaching users
  • Human in the Loop: Routing complex cases for human review
  • Adaptive Interventions: Customizing responses based on violation severity

Flow from Evaluate → Intervene​

Intervene closes the loop from observation to action. While Evaluate identifies risks via Laws—ranging from prompt injection to unauthorized data access—Intervene enforces policy decisions to mitigate or prevent harm.

Together, Evaluate and Intervene form a continuous monitoring-and-response system for LLM usage.

Example:​

  1. An Evaluator for detecting data exfiltration returns a high confidence score (>0.85) with supporting signals.
  2. A Law in Evaluate that calls the data exfiltration Evaluator decides that a user attempted to exfiltrate sensitive data via prompt injection.
  3. Intervene receives the violation event.
  4. A matching policy triggers a multi-step response: block the prompt, notify the compliance team, and log the event.

Why ThirdLaw Intervene?​

  • Execution Engine for Laws: Transforms detection signals into live, traceable action.
  • Designed for IT and Security Teams: Prioritizes governance, auditability, and operational control.
  • Extensible & Programmable: Adapts to both simple rule-based and complex multi-evaluator Laws.
  • Latency-Aware: Operates in real-time or batch mode as appropriate to the risk.
  • Feedback Loops: Captures data to improve Laws and future intervention strategies.