k_return is the strength of the spring that pulls internal log-state toward the declared normal x*. If you only remember one GSRF parameter, remember this one.
Spring on vs spring off
| Setting | Behavior | Evidence cue |
|---|---|---|
| k_return ≈ 0 | Tracker-like; step 50→65 settles near 65 | Matches EMA path on step identity table |
| Balanced k_return | Halfway settle ~55; peaks compressed | Public Practical identity |
| Reference (open attractor) | Hugs x*; weak live tracking | Open-loop twin — not a live tracker product |
The intentional trade-off
- Buy: mid-band oscillation kill, peak compression toward normal
- Pay: tracking MAD vs raw (0/52 wins in excellence hunt v2)
Mechanism ablations (workbench v3) treat k_return as the soul of Practical GSRF. Marketing that hides this trade is lying about the product.
Tuning posture (claim-safe)
Public evaluation uses balanced-style defaults on /try. Production presets (Pro trial) are for defined scopes—not “crank k_return to max for free lunch.”
Verdict
If you need the spring, you want GSRF. If you need zero spring cost, use EMA/Kalman. See What GSRF really is and peaks vs MAD.
Sources: mechanism v3 · characterization v4 step · EXCELLENCE_FINDING · glossary