diff --git a/docs/performance/2026-09-05-scroll-investigation.md b/docs/performance/2026-09-05-scroll-investigation.md
new file mode 100644
index 0000000..54bd2ee
--- /dev/null
+++ b/docs/performance/2026-09-05-scroll-investigation.md
@@ -0,0 +1,54 @@
+# Scrolling performance investigation
+
+Production baseline: `f3f2b03` on `codex/portfolio-performance`.
+
+Hardware: ASUS ROG Strix SCAR 16 G635LX, Intel Core Ultra 9 275HX,
+hardware-backed ANGLE Intel Mesa ARL graphics, Chrome 152. Viewport
+1440 × 900, DPR approximately 1, canvas 1425 × 900. Effective High
+and Medium tiers were recorded throughout each measurement. Phone testing
+is deferred.
+
+## Findings
+
+Matched 45-second High-quality scrolling runs after five seconds of warm-up:
+
+| Condition | Presented frames/second | p95 presentation interval |
+| --- | ---: | ---: |
+| Normal page | 71.4 | 25 ms |
+| Chapter masks disabled | 116.1 | 12.5 ms |
+
+The 3D scene, render resolution, bloom, and motion were unchanged. Disabling
+masks changes the appearance and was diagnostic only. Native wheel scrolling
+reproduced the pattern. A separate 32-second Medium comparison measured
+80.2 versus 142.4 animation callbacks/second, with p95 intervals of 20.9
+versus 12.4 ms.
+
+Full traces showed approximately 8,800 versus 840 page raster operations.
+Static masks performed substantially better than continually changing masks.
+Layer promotion, paint containment, and viewport-sized mask images did not
+resolve the slow frames. Whole-world GPU samples had a median of 2.7 ms;
+JavaScript scene submission was approximately 0.7 ms. Removing the entire
+car or terrain did not resolve the 25 ms p95 scrolling intervals. Bypassing
+bloom helped less than disabling masks and still exceeded the 60 FPS budget.
+
+The absolute rate in this session exceeded the earlier approximately 45 FPS
+control, including when rerunning the original measurement harness. The
+reason for that session difference was not isolated. These are matched
+same-session diagnostic comparisons, not newly shipped improvements.
+Detailed tracing adds overhead; GPU timer intervals may include stalls and
+cannot be added directly to displayed frame intervals.
+
+## Implementation direction
+
+Keep the fades visually close while separating the stationary reading edge
+from scrolling content, allowing rasterized content and mask data to be
+reused. Preserve native document scrolling, navigation, focus, content,
+3D quality, audio, and shared journey outputs. Validate responsive layouts,
+short viewports, reversals, quality changes, Static mode, reduced motion,
+resource cleanup, and sustained rendering before retaining the change.
+
+The user allows modest visual differences if the result does not look worse.
+No production changes were made during investigation. All browser runs
+inventoried and reused the single existing Chrome tab/window. Temporary
+scripts, traces, screenshots, and detailed measurements remain in the ignored
+`.perf/investigation/` folder; no tests or frameworks were added.