Monitoring drift into failure is hindered by Euclidean anomaly detection that can conflate safe operational trade-offs with risk accumulation in signals expressed as shares, and by architectural churn that makes fixed schemas (and learned models) stale before rare boundary events occur. Rasmussen's dynamic safety model motivates drift under competing pressures, but operationalizing it for software is difficult because many high-value operational signals (effort, remaining margin, incident impact) are compositional and their parts evolve. We propose a vision for drift observability on the simplex: model drift and boundary proximity in Aitchison geometry to obtain coordinate-invariant direction and distance-to-safety in interpretable balance coordinates. To remain comparable under churn, a monitor would continuously refresh its part inventory and policy-defined boundaries from engineering artifacts and apply lineage-aware aggregation. We outline early-warning diagnostics and falsifiable hypotheses for future evaluation.

翻译：监测故障漂移受到以下因素阻碍：欧几里得异常检测可能将安全操作权衡与以份额形式表达的信号中的风险累积相混淆，以及架构变动导致固定模式（及已学习模型）在罕见边界事件发生前就已过时。拉斯穆森的动态安全模型揭示了竞争压力下的漂移机制，但将其应用于软件系统却面临挑战，因为许多高价值操作信号（工作量、剩余裕度、事故影响）具有组合性且其组成部分持续演化。我们提出单纯形上漂移可观测性的理论框架：在艾奇逊几何中建模漂移与边界接近度，从而在可解释的平衡坐标系中获得坐标不变的安全方向与安全距离度量。为在架构变动下保持可比性，监测系统需持续从工程制品中更新其组件清单与策略定义边界，并应用谱系感知聚合方法。我们进一步构建了面向未来评估的早期预警诊断体系与可证伪假设。