When should I not use GSRF?

When you need best MAD tracking, multi-minute delayed command smoothing, fixed-FA residual detection, or a universal always-beats-EMA filter.

What is the safety cage?

The published set of non-claims and red lights that marketing may not invert — tracker, delay, fixed-FA, universal win.

Is the cage the same as certification?

No. It is product honesty for procurement, not a SIL or regulatory certificate.

When not to use GSRF (the safety cage)

Trust is structural when vendors publish red lights with the same energy as green lights. Below is the public safety cage for Zero Overshoot / GSRF Practical.

Red — do not use for these jobs

1. Delayed closed-loop command smoothing

MAD to delayed measured plant vs EMA (phase4-style delay packs): +59% worse at τ=3 min, +97% at 6 min, +106% at 10 min. The spring fights delayed truth. Use other tools for that path.

2. Fixed false-alarm event detection

At matched FA rate on synthetic excursions (residual mode): GSRF recall ~0.44 vs EMA ~0.83. Peak compression and higher normal residual raise the calibrated threshold. Pair a detector with GSRF—or use EMA residual for that task.

3. Minimum MAD tracking vs raw

Excellence hunt v2: 0 / 52 wins vs best EMA on tracking MAD. If your KPI is hug-the-raw, turn the spring down (or use EMA)—and accept losing the niche.

Green — designed for these jobs

Mid-band oscillation kill (~+35% osc31; ~−70% sweet-band spectral power on locked packs)
Peak compression toward defined normal (~38% mean; 49/52 wins)
Adaptive x* when the plant’s normal truly moves
Pre-model / pre-AI signal hygiene (calm first; detect separately)

Procurement wording

Prefer: “soft thermostat with published band and failure modes.” Avoid: “universal optimal filter” or “eliminates false alarms.” See the claim ladder on the Workbench page.

Sources: Evidence §4 · characterization v4 delay · probes v5 alarm · excellence hunt v2. /evidence#boundaries · /workbench