Streaming for free: one weight vector for the filter and the smoother

A two-stage sequence regressor that must run as a causal filter while data arrive and as a smoother once the window is complete leaks if the corrector is allowed to read the label during training. It learns to copy neighbours, and the two modes diverge at deployment. We train the corrector on the model's own pointwise draft only, and draw a causal or bidirectional mask per batch with equal probability. Two short propositions bound the gap (label leakage is a covariate shift bounded by the leaked fraction times the draft error; the mask draw puts both modes in one objective). The ordering of protocols is the same on oil temperature, air quality and ECG as on the aerodynamic case study where we found it. On ETTh1 a 4k-parameter model with one lag of the target and the covariates reaches $0.65\^\circ$C in both modes, where zero-shot Chronos, TimesFM and TinyTimeMixer given 512 hours of the target reach $0.75$-$0.81$. The case study is a $5{,}285$-parameter polar predictor that streams at $81\\mu$s per 68-point window on one CPU thread.