diff --git a/README.md b/README.md
index 1031214..3c51db4 100644
--- a/README.md
+++ b/README.md
@@ -120,6 +120,30 @@ hour, at which point the alarm is the outage.
 `smartctl -H` across two dozen drives runs hourly, not per loop. Absolute and trend
 checks stay on the fast loop.
 
+### Ramp width controls how *bouncy* the fans are
+
+Changing fan noise is more irritating than constant fan noise, and ramp width is the lever
+for it. The span sets the sensitivity: 22 points of fan over a 7c ramp is 3.14%/degC, so
+1.6c of CPU drift is enough to cross a 5% deadband. The same 22 points over 11c is
+2%/degC, needing 2.5c.
+
+Replaying 180 recorded samples from `iz-pve0` under load (CPU 57-69c) through candidate
+curves, counting how often the fan would actually move:
+
+| Ramp | Deadband | Fan changes | Distinct speeds |
+|---|---|---|---|
+| 62-69c | 5 | 19 | 28,31,34,37,40,43,46,50 |
+| 62-69c | 8 | 6 | 28,37,46,50 |
+| 62-73c | 5 | 5 | 28,34,40 |
+| **62-73c** | **8** | **1** | 28,36 |
+
+Confirmed live against a multi-hour zpool migration: **19 fan changes became 2** over a
+comparable window, average fan speed dropped slightly (40% to 38%), and **CPU temperature
+did not move** (64-68c against 64-69c before). The extra airflow had been buying nothing.
+
+Widen the ramp and raise the deadband together. The cost is lag - a deadband of 8 at
+2%/degC is about 4c of slack - which is affordable when you are 8-14c below any threshold.
+
 ## Calibrating a host
 
 **The defaults are not portable and the fans will be wrong on an uncalibrated host.**