If you score filters only on MAD-to-raw, GSRF Practical will look “wrong.” In excellence hunt v2 it recorded 0 / 52 wins on tracking MAD vs best EMA. That is not a secret failure—it is the spring.
The step test that explains the product
Step the raw signal from 50 → 65 with x* frozen at the pre-step normal (log-space):
- EMA settles at 65 (tracks the new level).
- GSRF k_return = 0 settles at 65 (spring off ≈ tracker-like).
- GSRF Practical balanced settles near ~55 (halfway home).
- GSRF Reference stays near 50 (open attractor twin).
Balanced GSRF equilibrates observation pull against the restoring spring to x*. That equilibrium is peak compression toward normal.
Locked peak results
Across characterization families, peak-deviation vs best EMA showed about ~38% mean advantage and 49 / 52 dataset wins. Use that language—not “zero overshoot on every plant.”
If you need full settle after a permanent jump
Move x* (adaptive thermostat). Do not demand that balanced GSRF track a new permanent level while keeping the spring at full strength. See the adaptive note.
Sources: Evidence §1 identity · characterization v4 step_up · excellence hunt v2 peak_dev. /evidence#identity