diff --git a/frame/perf/session-h18/README.md b/frame/perf/session-h18/README.md
new file mode 100644
index 0000000..d114493
--- /dev/null
+++ b/frame/perf/session-h18/README.md
@@ -0,0 +1,17 @@
+# Headset session 18: the end-delay band at 0.5 ms steps (4.0 again, 4.5, 3.5)
+
+Same build as sessions 16 and 17. Owner: "all of them had doubling in the right eye (launch 3, 3.5 ms,
+maybe the least but still very present) and super slight in the left, especially in the water and visible
+on the HUD."
+
+Logs: ends tight in every launch (4000: 4-4.5 ms 4635 of 5743, 4.5-5 ms 838; 4500: 4.5-5 ms 5100 of 6335;
+3500: 3.5-4 ms 7719 of 11951, 4-4.5 ms 2841). Session 16's clean 4000 launch had the same distribution
+(4-4.5 ms 5821 of 7324, 4.5-5 ms 1262) and the same locate-to-display offset (43.7-43.8 ms p50 in all).
+Nothing the app logs separates the clean 4000 launch from the doubled one.
+
+Reading: the clean band is not a fixed offset from the app's wake-up; the relation between the app's frame
+clock (xrWaitFrame) and the point in the refresh where the compositor takes each eye's image differs per
+launch and drifts (session 17's 2 s beat). Ending later than the wake-up is the right lever (session 16),
+but the end must be timed against the compositor's own clock, which the app cannot see today. Next: lane W
+makes that phase observable (per-frame compositor timing beside the app's wake-up and end times) and
+designs the lock.