传统组态看现场，AI 原生组态预演现场。
Traditional mimic shows the field. AI-native mimic pre-simulates it.
左边是你熟悉的工艺图：系统先给建议，拖一下阀门或泵，图上每个设备的「现值 → 预测值」实时变；点「▶ 推演」就看着供水温度顺着真实轨迹涨到达标。右边是工程控制面板：同一份动作的 7 项指标、安全判决、AI 解释——运行人员看图点执行，技术员看右边对数。两边同一个引擎、同一份状态。
Left is the mimic you already read: the system suggests first, drag the valve/pump and every component updates live 'now → predicted'; hit ▶ to watch the supply temp climb along the real trajectory to setpoint. Right is the engineering panel: the same action's 7 metrics, safety verdict, AI explanation. Operators act on the diagram; engineers check the numbers on the right — one engine, one shared state.
- 在安全边界内（Δ 3.0%）。Within safety bounds (Δ 3.0%).
▸经验校准闭环 · 工程（执行→实际→残差→自校准）Online calibration loop · engineering (execute→actual→residual→self-calibrate)
执行后，实际效果回写、与预测比对，残差驱动模型自校准。这里换热器的「真实」效率是 0.63，但模型初值是 0.70——它并不知道。每执行一次，模型从残差里修一点 eps_hx，几次之后就逼近真实、残差归零。这正是把一次性参数辨识变成持续校准，也是边缘场景库的数据来源。
After execution, the actual outcome is logged, compared to the prediction, and the residual self-calibrates the model. The HX's true effectiveness is 0.63, but the model starts believing 0.70 — it doesn't know. Each execution nudges eps_hx from the residual; after a few it converges and the residual collapses. This turns one-shot identification into living calibration, and is the data source for the edge-case library.
每条记录通过 /api/decisions 写入决策库（未配置 DATABASE_URL 时仅在浏览器内运行，演示不受影响）。这些「状态/预测/实际/残差」记录同时是经验校准的燃料和世界模型的带标签边缘场景。
Each cycle is POSTed to /api/decisions (no-op without DATABASE_URL — the demo runs in-browser regardless). These {state, prediction, actual, residual} records are both the calibration fuel and the world model's labeled edge cases.