TELOS SB 243 Child Safety Validation Dataset

TELOS SB 243 Child Safety Validation Dataset Reproducible validation artifacts for TELOS geometric governance against California SB 243 child safety harm categories. Key Results Metric Value Attack Success Rate 0.00% False Positive Rate 74.00% Violation Defense Rate 100.00% Contents Test Data sb243_data/sb243_test_attacks.csv - 50 harmful attack prompts (suicide, self-harm, sexual content, eating disorders) sb243_data/sb243_benign_contrastive.csv - 50 benign helper/educational queries Configuration config/sb243_child_safety_pa_config.json - Primacy Attractor definition for SB 243 compliance Validation Scripts run_sb243_validation.py - Attack validation (measures ASR) run_false_positive_validation.py - False positive validation (measures FPR) Results sb243_validation_results.json - Detailed attack validation results sb243_false_positive_results.json - Detailed FPR results Reproduction pip install sentence-transformers numpy# Run attack validationpython run_sb243_validation.py# Run false positive validationpython run_false_positive_validation.py Embedding Model All validation uses sentence-transformers/all-MiniLM-L6-v2 (384 dimensions). Governance Thresholds Tier 1 (Autonomous Block): fidelity >= 0.25 Tier 2 (Review Block): fidelity in [0.15, 0.25) Tier 3 (Expert Escalation): fidelity < 0.15 Regulatory Context California SB 243 (Chapter 883, Statutes of 2024) - Child safety requirements for social media platforms. Citation If you use this dataset, please cite the TELOS project: GitHub: https://github.com/TelosSteward/TELOS License Apache 2.0 Validation Date: 2024-12-22