按输出合同、失败代价、时限和运行轨迹路由模型，用每个完成任务的总成本取代 Token 单价。
从触发到合并后稳定运行，分开测量 Routine 健康、任务验收、CI、审查和工作流流失。
AI 安全控制需要独立日志、合成探针、自动到期的例外和端到端证据，避免关闭拦截时同时失去监测。
Route LLM tasks by output contract, failure cost, deadline, and trajectory signals, then optimize cost per completed task.
Measure Claude Code routines from trigger to unreverted merge, separating run health, task success, CI, review, and workflow losses.
AI safety controls need independent logs, synthetic probes, expiring exceptions, and end-to-end evidence when blocking is disabled.
本地视觉与云端推理分层，可以让家庭图像留在设备内，再用校准升级、事件日志和可逆动作建立闭环。
A local-vision, cloud-reasoning architecture keeps home images private while using calibrated escalation, event logs, and reversible actions.
WhatsApp Scam Alert 用端侧推理、机密分析、透明账本和失败关闭验证，在端到端加密下建立可审计的诈骗识别链路。
WhatsApp Scam Alert shows how on-device ML, confidential analytics, transparency logs, and fail-closed verification can protect E2EE messaging.
