L3)¶

CT Toolkit uses a three-tier escalation model. Each tier only runs when the previous one detects a potential issue — keeping costs and latency near zero for healthy agents.

L1 — Embedding Cosine Similarity (ECS)¶

Runs on every call. Zero extra API cost.

divergence = 1.0 - cosine_similarity(response_vector, reference_vector)

L2 — LLM-as-Judge¶

Triggered when L1 exceeds l2_threshold. An independent model evaluates the response against the kernel rules.

L3 — ICM Probe Battery¶

Triggered when L2 returns misaligned or L1 exceeds l3_threshold. Runs a battery of ethical probe scenarios.