With a frozen normal, GSRF will not fully adopt a permanent level change— that is peak-compression identity. Production plants still change operating points. The fix is not to lie about the spring; it is to move the thermostat slowly.
Synthetic probe (50 → 70 → 40)
Mean absolute mean-error vs true regime base on the shifted regimes (B+C):
- GSRF frozen prefix x*: ~9.7
- GSRF fixed full-series x*: ~9.7
- GSRF adaptive (window 200): ~1.5
- EMA_low: ~0.5 (still the best pure tracker)
In the original regime, adaptive GSRF still beat EMA on oscillation (osc31 ~0.50 vs ~0.78 in that probe).
Real-data nuance
On the realdata v6 matrix, adaptive often keeps phase5/phase6.2 osc wins and can help multi-regime MBP-style traces. It can hurt flat SpO2-like series—do not enable adaptive blindly on clipped nearly-constant signals.
Implementation sketch
Causal adaptive x*: trailing median of log(signal) over a window, updated online while the same GSRF Practical state continues. That is what the workbench probe used—not a second filter fighting the first.
Sources: Evidence §3 · probes v5 adaptive · realdata v6. /evidence#adaptive