OpenAI 前沿治理框架：当模型能力撞上监管红线

"""
governance_gate.py — 最小前沿模型部署治理门禁

使用方式：
  python governance_gate.py --model gpt-next --capability-score 0.82 --security-tier 2

假设：
  - capability_score: 0-1 浮点数，来自红队评估流程
  - security_tier: 1-3 整数，1=基础, 2=增强, 3=最高
  - eu_systemic_risk: 模型训练成本是否超过 EU AI Act 系统性风险阈值
  - california_threshold: 训练成本是否超过加州法案阈值

运行前无需额外依赖，仅用标准库。
"""

import argparse
import json
import sys
from datetime import datetime, timezone


# ── 治理规则配置 ──────────────────────────────────────────────

RULES = {
    # 能力阈值：超过此分数必须触发额外审查
    "capability_threshold": 0.75,
    # 安全等级要求：能力超过阈值时，security_tier 必须 >= 此值
    "min_security_tier_for_high_capability": 3,
    # EU 系统性风险模型必须完成红队报告
    "eu_requires_redteam_report": True,
    # 加州阈值模型必须有关停机制文档
    "california_requires_shutdown_doc": True,
}


def evaluate_gate(
    model: str,
    capability_score: float,
    security_tier: int,
    eu_systemic_risk: bool,
    california_threshold: bool,
    redteam_report_exists: bool,
    shutdown_doc_exists: bool,
) -> dict:
    """评估治理门禁，返回决策和阻塞项。"""

    blockers = []
    warnings = []

    # 规则 1：能力超过阈值 → 安全等级必须达标
    if capability_score >= RULES["capability_threshold"]:
        if security_tier < RULES["min_security_tier_for_high_capability"]:
            blockers.append(
                f"能力分数 {capability_score:.2f} 超过阈值 "
                f"{RULES['capability_threshold']}, 但安全等级仅 {security_tier}，"
                f"需要 >= {RULES['min_security_tier_for_high_capability']}"
            )
        else:
            warnings.append(
                f"能力分数 {capability_score:.2f} 超过阈值，安全等级 {security_tier} 已达标"
            )

    # 规则 2：EU 系统性风险 → 必须有红队报告
    if eu_systemic_risk and not redteam_report_exists:
        blockers.append("EU AI Act 系统性风险模型，缺少红队评估报告")

    # 规则 3：加州阈值 → 必须有关停机制文档
    if california_threshold and not shutdown_doc_exists:
        blockers.append("加州法案阈值模型，缺少关停机制文档")

    decision = "BLOCK" if blockers else ("PROCEED_WITH_WARNINGS" if warnings else "PROCEED")

    result = {
        "timestamp": datetime.now(timezone.utc).isoformat(),
        "model": model,
        "capability_score": capability_score,
        "security_tier": security_tier,
        "eu_systemic_risk": eu_systemic_risk,
        "california_threshold": california_threshold,
        "decision": decision,
        "blockers": blockers,
        "warnings": warnings,
    }

    return result


def main():
    parser = argparse.ArgumentParser(description="前沿模型部署治理门禁")
    parser.add_argument("--model", required=True, help="模型标识符")
    parser.add_argument("--capability-score", type=float, required=True, help="能力评估分数 0-1")
    parser.add_argument("--security-tier", type=int, choices=[1, 2, 3], required=True, help="安全等级 1-3")
    parser.add_argument("--eu-systemic-risk", action="store_true", help="是否为 EU 系统性风险模型")
    parser.add_argument("--california-threshold", action="store_true", help="是否超过加州法案训练成本阈值")
    parser.add_argument("--redteam-report", action="store_true", help="红队评估报告是否存在")
    parser.add_argument("--shutdown-doc", action="store_true", help="关停机制文档是否存在")
    args = parser.parse_args()

    result = evaluate_gate(
        model=args.model,
        capability_score=args.capability_score,
        security_tier=args.security_tier,
        eu_systemic_risk=args.eu_systemic_risk,
        california_threshold=args.california_threshold,
        redteam_report_exists=args.redteam_report,
        shutdown_doc_exists=args.shutdown_doc,
    )

    print(json.dumps(result, indent=2, ensure_ascii=False))

    if result["decision"] == "BLOCK":
        print("\n⛔ 部署被阻塞，请解决以上 blockers 后重新评估。", file=sys.stderr)
        sys.exit(1)
    elif result["decision"] == "PROCEED_WITH_WARNINGS":
        print("\n⚠️  部署可继续，但请注意 warnings。", file=sys.stderr)
    else:
        print("\n✅ 部署门禁通过。", file=sys.stderr)


if __name__ == "__main__":
    main()

# 能力分数 0.82 超阈值，但安全等级只有 2，且缺少红队报告
python governance_gate.py \
  --model gpt-next \
  --capability-score 0.82 \
  --security-tier 2 \
  --eu-systemic-risk \
  --california-threshold

# 输出：
# {
#   "decision": "BLOCK",
#   "blockers": [
#     "能力分数 0.82 超过阈值 0.75, 但安全等级仅 2，需要 >= 3",
#     "EU AI Act 系统性风险模型，缺少红队评估报告",
#     "加州法案阈值模型，缺少关停机制文档"
#   ]
# }

# 所有条件满足
python governance_gate.py \
  --model gpt-next \
  --capability-score 0.82 \
  --security-tier 3 \
  --eu-systemic-risk \
  --california-threshold \
  --redteam-report \
  --shutdown-doc

# 输出：
# {
#   "decision": "PROCEED_WITH_WARNINGS",
#   "warnings": ["能力分数 0.82 超过阈值，安全等级 3 已达标"]
# }

# governance_rules.yaml
capability_threshold: 0.75
min_security_tier_for_high_capability: 3
eu_requires_redteam_report: true
california_requires_shutdown_doc: true
# 可扩展：自定义伤害场景阈值
harm_scenarios:
  biological_attack_path:
    threshold: 0.60
    required_action: restrict_deployment
  mass_manipulation:
    threshold: 0.70
    required_action: enhanced_monitoring