diff --git a/PHASE7_SQUIRREL_AB.md b/PHASE7_SQUIRREL_AB.md
new file mode 100644
index 00000000..c49c64bf
--- /dev/null
+++ b/PHASE7_SQUIRREL_AB.md
@@ -0,0 +1,53 @@
+# Phase 7: Squirrel vs NSIS A/B
+
+NSIS is the shipped Windows installer and **stays the default**. Squirrel is a
+candidate only — it wins, and replaces NSIS, **only if** it is measurably faster
+*and* its auto-update rollback works on real Win 10/11 machines. Otherwise NSIS
+stays. This doc is the procedure to make that call with data, not vibes.
+
+## Build both from the same commit
+
+```powershell
+# NSIS (default, what ships today)
+pwsh scripts\build-app-win.ps1 -Sign        # -> electron\dist\OpenSwarm-Setup-x64.exe
+
+# Squirrel (candidate), same staged tree / same SHA
+pwsh scripts\build-app-win.ps1 -Sign -Squirrel
+```
+
+The `-Squirrel` switch only overrides `win.target` (via
+`--config.win.target=squirrel`); signing, extraResources, and the bundled
+python/node/router are identical, so any measured difference is the installer
+itself, not the payload. Confirm both report the same provenance sha (Settings
+-> About -> Build, or the `[provenance]` line in backend.log).
+
+## Measure on REAL Windows 10 and 11 (x64), clean machines
+
+For each installer, on a fresh VM/box (no prior OpenSwarm install):
+
+| Metric | How |
+|---|---|
+| Install time | wall-clock from launching the installer to the app window appearing |
+| First paint | `[perf] first-paint` in backend.log (`scripts/perf/parse-timing.js`) |
+| Backend ready | `[perf] backend-http-ready` |
+| Crashes | any crash on first launch; check `%APPDATA%\OpenSwarm\Crashpad` |
+| Auto-update | install an older build, then this one; confirm it detects, downloads, installs on quit, relaunches on the new version |
+| Rollback | after an update, force a downgrade/rollback path; confirm the previous version comes back cleanly and the feed isn't corrupted |
+
+Run each 3x per OS and take the median. Compare against the Phase 0 baseline
+(file count 11,247 / 1.2 GB) and NSIS's own numbers.
+
+## Decision gate
+
+- Squirrel **wins** only if: median install + first-paint + backend-ready are
+  faster than NSIS on BOTH Win 10 and Win 11, AND auto-update works, AND rollback
+  works (including rebuilding whatever feed Squirrel's differential updates need).
+- Any of those fail -> **NSIS stays**, revert the target, keep `-Squirrel` as a
+  dead experiment flag or remove it.
+
+## Why this is the last phase
+
+Squirrel changes the update feed format and rollback semantics. Switching it in
+without the rollback feed rebuilt strands users on a broken updater — the exact
+failure the rest of this plan exists to prevent. So it goes last, behind a flag,
+and only on proof.