CDRA: Constraint-based Dynamic Receptivity Architecture

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.

Validation

Metric	Result
Models tested	16 instances, 6 architectures, 8 model families
Unsolicited advice (baseline)	100% on emotional inputs
Unsolicited advice (CDRA)	0% on emotional inputs
Task completion (CDRA)	100% (unchanged)
Cross-architecture consistency	Full (MoE, Dense, GPT, Gemini, Qwen, Llama)
Minimum effective constraint	~70 words
Stability (10-round stress test)	0% drift
Blind third-party validation	Cohen's κ ≈ 1.0

Files

• CDRA_English_Paper_v2_20260628.pdf — Formal paper
• CDRA_Experimental_Report_EN_20260628.pdf — Experimental report
• CDRA_Graded_Constraints_EN_20260628.pdf — Graded constraint library
• CDRA_MVP_Behavior_OS_EN_20260628.pdf — Minimum viable behavior OS
• CDRA_vs_Prompt_Engineering_EN_20260628.pdf — Comparison with prompt engineering
• README_Zenodo_Package_20260628.pdf — Package guide
• replication/ — One-click replication kit (run_cdra.py)

Citation

@misc{xiao2026cdra,
  title={CDRA: Constraint-based Dynamic Receptivity Architecture},
  author={Xiao, Han},
  year={2026},
  howpublished={Zenodo preprint},
  doi={10.5281/zenodo.20993162}
}

Related work

• Apert: 10.5281/zenodo.21005888
• Reception Science: 10.5281/zenodo.21078023
• Unloading Hypothesis: 10.5281/zenodo.21101755

License: CC-BY 4.0