
Most LLM evaluation repos break when a single-vendor judge agrees with its own hallucinations or an unrouted frontier model burns $40/day on intake classification.
The Human Creative Enterprise AI Eval & REA Pro Kit packages four dependency-light Python & Node reference systems with source code, JSONL datasets, SHA-256 verification receipts, and a 300-DPI Executive Benchmark PDF.
Human_Creative_AI_Eval_and_REA_Pro_Kit.zip)01_judge_trust_4vendor_harness/: Injects defect classes into golden items and scores cross-family judge panels (generate.py, inject.py, judge.py, score.py, verify_clean.py) to measure judge self-preference bias.02_six_rung_cost_router/: Enterprise 6-rung model-ladder cost router across intake, policy validation, and cross-family verification with JSONL audit receipts (six_rung_cost_router.py + Playbook).03_50_vector_guardrail_audit_suite/: 50-vector behavioral red-team runner (run_50_vector_guardrail_audit.py) plus executive HTML/PDF report templates (sample_ai_audit_report.pdf).04_rea_har_to_agent_env/: Deterministically compiles browser .har traces into reproducible offline agent eval environments + 300-DPI executive_benchmark.pdf.Commercial Developer License: Perpetual single-seat license to adapt, embed, and white-label.