Control teams should treat GSRF as a soft thermostat on measurement or supervisory command shaping — with a mandatory shadow phase — not as a magic smoother that always improves closed-loop error.
Recommended adoption pattern
- Shadow: run GSRF offline/parallel; compare osc31, peaks, MAD vs EMA on recorded loops
- Measure delay: if command→measured delay is multi-minute class, read the red light before any live command path
- Shape carefully: prefer measurement hygiene or supervisory trim around a declared normal
- Gate production: commercial license + your acceptance tests
Locked delay red light
On locked packs, MAD-to-measured under balanced Practical was about +59% / +97% / +106% worse than EMA at 3 / 6 / 10 minute delays. That is why we refuse “drop-in delayed plant feedback command filter” marketing.
Where it still helps control-adjacent stacks
- Calming oscillatory measured PV before monitoring / soft-sensing
- Peak compression toward an operating normal for wear/chatter metrics
- Industrial actuator-style batteries (synthetic Phase 5; DAMADICS partial/loss published)