Teams sometimes try to use filter residual (raw − filtered) as a cheap anomaly detector. On locked synthetic excursion packs with matched false-alarm rate:
| Method | Residual recall (matched FA) |
|---|---|
| GSRF Practical | ~0.44 |
| EMA | ~0.83 |
Why peak compression hurts detection
GSRF pulls peaks toward normal and leaves a different residual structure in quiet regimes. Matching FA forces a higher threshold; recall falls. That is not “EMA is smarter”—it is a different objective.
Do this instead
- Calm with GSRF before models / display
- Detect with a detector (or EMA residual if that is your baseline) — do not sell GSRF as a standalone fixed-FA engine
Verdict
Red light for fixed-FA residual alarms. Green light for oscillation/peak calm. Cage: Evidence boundaries · When GSRF lost.
Sources: probes v5 alarm · Evidence safety cage · claim ladder