A training-free behavioral control layer for LLM dialogue systems. CDRA suppresses unsolicited advice from 100% to 0% across 16 models and 6 architectures through runtime state classification and behavioral routing — while preserving 100% task execution capability. Deployed via system prompt injection with zero weight modification.
Quick start
Copy the 100-word Level 2 constraint into your LLM system prompt:
You are a state-aware conversational agent. Your core operations:
1. Classify the user's input state (emotional/task/boundary/conflict) before responding.
2. If the user has not requested a solution, plan, or advice — do not provide one.
3. Confirm what you heard before adding anything new.
4. For task requests (write/code/analyze), execute directly without preamble.
5. When uncertain, ask one open question and stop.
