diff --git a/docs/re/data/sweep-leaf-ramp.txt b/docs/re/data/sweep-leaf-ramp.txt
new file mode 100644
index 00000000..c8b8914d
--- /dev/null
+++ b/docs/re/data/sweep-leaf-ramp.txt
@@ -0,0 +1,64 @@
+# The sweep strips: the ramp is DECODED on the disc, and the GPU agrees.
+# 2026-08-31. sylpheed-port asked for "vertex alpha as a function of sweep
+# position". It is not a runtime curve to be sampled -- it is four keyframes.
+#
+# cargo run -p sylpheed-formats --example sweep_leaf_ramp -- GP_TITLE 5
+# tools/re-capture/sweep_positions.py <draw logs>
+
+=== GP_TITLE entry 5 — ptloop01.rat (leaf 284 bytes) ===
+    declared loop length: 600 units
+  pteff03.t32 — 4 keyframes
+    kf0  t=0     x=-639   y=270    sx=100  sy=600  rot=30    fade=FFFFFFFF (alpha 255)
+    kf1  t=150   x=-39    y=270    sx=100  sy=600  rot=30    fade=80FFFFFF (alpha 128)
+    kf2  t=540   x=1521   y=270    sx=100  sy=600  rot=30    fade=FFFFFFFF (alpha 255)
+    kf3  t=600   x=1521   y=270    sx=100  sy=600  rot=30    fade=FFFFFFFF (alpha 255)
+
+=== GP_TITLE entry 5 — ptloop02.rat (leaf 283 bytes) ===
+    declared loop length: 720 units
+  pteff03a.t32 — 4 keyframes
+    kf0  t=0     x=1721   y=270    sx=100  sy=800  rot=-45   fade=00FFFFFF (alpha   0)
+    kf1  t=150   x=1111   y=270    sx=100  sy=800  rot=-45   fade=80FFFFFF (alpha 128)
+    kf2  t=630   x=-839   y=270    sx=100  sy=800  rot=-45   fade=FFFFFFFF (alpha 255)
+    kf3  t=720   x=-839   y=270    sx=100  sy=800  rot=-45   fade=FFFFFFFF (alpha 255)
+
+DECLARED SLOPES, from the second segment of each ramp:
+  pteff03    x  -39 -> 1521 over t 150..540, alpha 128 -> 255   d(alpha)/dx = +0.0814
+  pteff03a   x 1111 -> -839 over t 150..630, alpha 128 -> 255   d(alpha)/dx = -0.0651
+
+MEASURED OFF THE GPU, three independent sessions:
+
+alpha vs position, per tall strip (design px, alpha level):
+  h=1134  n=3  x 442..486 px  alpha 239..242  d(alpha)/dx +0.0670
+  h=1303  n=3  x -435..-390 px  alpha 157..154  d(alpha)/dx -0.0670
+
+--
+alpha vs position, per tall strip (design px, alpha level):
+  h=1134  n=3  x -832..-819 px  alpha 135..136  d(alpha)/dx +0.0781
+  h=1303  n=3  x 403..416 px  alpha 47..45  d(alpha)/dx -0.1562
+
+--
+alpha vs position, per tall strip (design px, alpha level):
+  h=1134  n=4  x 237..307 px  alpha 222..227  d(alpha)/dx +0.0710
+  h=1303  n=4  x -1229..-1158 px  alpha 209..204  d(alpha)/dx -0.0710
+
+
+WHAT AGREES
+  * sign, both strips, every session: the 1134-tall strip moves +x with alpha
+    rising, the 1303-tall one moves -x with alpha rising. That is the declared
+    direction for pteff03 and pteff03a respectively.
+  * magnitude, within ~15 %: +0.067/+0.078/+0.071 against a declared +0.0814,
+    and -0.067/-0.071 against a declared -0.0651.
+  * the RANGE is exercised. Across sessions the measured alpha spans 45..242,
+    and 45 is below 128. Only pteff03a declares alpha below 128 -- pteff03's
+    ramp never leaves [128,255]. The strip measured at 45 is the 1303-tall one,
+    which the AABB geometry independently says is pteff03a. Two identifications,
+    one from height and one from alpha, agreeing.
+
+WHAT THIS CANNOT DO
+  * separate the two declared slopes. They are 25 % apart; NDC prints to two
+    decimals so one frame's dx is quantised to 6.4 px and alpha to one level,
+    and at 3-4 frames per session the noise is larger than the difference. The
+    -0.1562 row is that noise: 13 px of travel and 2 alpha levels.
+  * check ABSOLUTE phase. The log gives a rotated quad's AABB, and the mapping
+    from AABB-left to the element's declared x under rotation and pivot is not
+    established here. Only rates and ranges are compared.